source: branches/minix3-book/servers/pm/signal.c@ 9

/* This file handles signals, which are asynchronous events and are generally
 * a messy and unpleasant business.  Signals can be generated by the KILL
 * system call, or from the keyboard (SIGINT) or from the clock (SIGALRM).
 * In all cases control eventually passes to check_sig() to see which processes
 * can be signaled.  The actual signaling is done by sig_proc().
 *
 * The entry points into this file are:
 *   do_sigaction:   perform the SIGACTION system call
 *   do_sigpending:  perform the SIGPENDING system call
 *   do_sigprocmask: perform the SIGPROCMASK system call
 *   do_sigreturn:   perform the SIGRETURN system call
 *   do_sigsuspend:  perform the SIGSUSPEND system call
 *   do_kill:   perform the KILL system call
 *   do_alarm:  perform the ALARM system call by calling set_alarm()
 *   set_alarm: tell the clock task to start or stop a timer
 *   do_pause:  perform the PAUSE system call
 *   ksig_pending: the kernel notified about pending signals
 *   sig_proc:  interrupt or terminate a signaled process
 *   check_sig: check which processes to signal with sig_proc()
 *   check_pending:  check if a pending signal can now be delivered
 */

#include "pm.h"
#include <sys/stat.h>
#include <sys/ptrace.h>
#include <minix/callnr.h>
#include <minix/com.h>
#include <signal.h>
#include <sys/sigcontext.h>
#include <string.h>
#include "mproc.h"
#include "param.h"

#define CORE_MODE       0777    /* mode to use on core image files */
#define DUMPED          0200    /* bit set in status when core dumped */

FORWARD _PROTOTYPE( void dump_core, (struct mproc *rmp)                 );
FORWARD _PROTOTYPE( void unpause, (int pro)                             );
FORWARD _PROTOTYPE( void handle_sig, (int proc_nr, sigset_t sig_map)    );
FORWARD _PROTOTYPE( void cause_sigalrm, (struct timer *tp)              );

/*===========================================================================*
 *                              do_sigaction                                 *
 *===========================================================================*/
PUBLIC int do_sigaction()
{
  int r;
  struct sigaction svec;
  struct sigaction *svp;

  if (m_in.sig_nr == SIGKILL) return(OK);
  if (m_in.sig_nr < 1 || m_in.sig_nr > _NSIG) return (EINVAL);
  svp = &mp->mp_sigact[m_in.sig_nr];
  if ((struct sigaction *) m_in.sig_osa != (struct sigaction *) NULL) {
        r = sys_datacopy(PM_PROC_NR,(vir_bytes) svp,
                who, (vir_bytes) m_in.sig_osa, (phys_bytes) sizeof(svec));
        if (r != OK) return(r);
  }

  if ((struct sigaction *) m_in.sig_nsa == (struct sigaction *) NULL) 
        return(OK);

  /* Read in the sigaction structure. */
  r = sys_datacopy(who, (vir_bytes) m_in.sig_nsa,
                PM_PROC_NR, (vir_bytes) &svec, (phys_bytes) sizeof(svec));
  if (r != OK) return(r);

  if (svec.sa_handler == SIG_IGN) {
        sigaddset(&mp->mp_ignore, m_in.sig_nr);
        sigdelset(&mp->mp_sigpending, m_in.sig_nr);
        sigdelset(&mp->mp_catch, m_in.sig_nr);
        sigdelset(&mp->mp_sig2mess, m_in.sig_nr);
  } else if (svec.sa_handler == SIG_DFL) {
        sigdelset(&mp->mp_ignore, m_in.sig_nr);
        sigdelset(&mp->mp_catch, m_in.sig_nr);
        sigdelset(&mp->mp_sig2mess, m_in.sig_nr);
  } else if (svec.sa_handler == SIG_MESS) {
        if (! (mp->mp_flags & PRIV_PROC)) return(EPERM);
        sigdelset(&mp->mp_ignore, m_in.sig_nr);
        sigaddset(&mp->mp_sig2mess, m_in.sig_nr);
        sigdelset(&mp->mp_catch, m_in.sig_nr);
  } else {
        sigdelset(&mp->mp_ignore, m_in.sig_nr);
        sigaddset(&mp->mp_catch, m_in.sig_nr);
        sigdelset(&mp->mp_sig2mess, m_in.sig_nr);
  }
  mp->mp_sigact[m_in.sig_nr].sa_handler = svec.sa_handler;
  sigdelset(&svec.sa_mask, SIGKILL);
  mp->mp_sigact[m_in.sig_nr].sa_mask = svec.sa_mask;
  mp->mp_sigact[m_in.sig_nr].sa_flags = svec.sa_flags;
  mp->mp_sigreturn = (vir_bytes) m_in.sig_ret;
  return(OK);
}

/*===========================================================================*
 *                              do_sigpending                                *
 *===========================================================================*/
PUBLIC int do_sigpending()
{
  mp->mp_reply.reply_mask = (long) mp->mp_sigpending;
  return OK;
}

/*===========================================================================*
 *                              do_sigprocmask                               *
 *===========================================================================*/
PUBLIC int do_sigprocmask()
{
/* Note that the library interface passes the actual mask in sigmask_set,
 * not a pointer to the mask, in order to save a copy.  Similarly,
 * the old mask is placed in the return message which the library
 * interface copies (if requested) to the user specified address.
 *
 * The library interface must set SIG_INQUIRE if the 'act' argument
 * is NULL.
 */

  int i;

  mp->mp_reply.reply_mask = (long) mp->mp_sigmask;

  switch (m_in.sig_how) {
      case SIG_BLOCK:
        sigdelset((sigset_t *)&m_in.sig_set, SIGKILL);
        for (i = 1; i <= _NSIG; i++) {
                if (sigismember((sigset_t *)&m_in.sig_set, i))
                        sigaddset(&mp->mp_sigmask, i);
        }
        break;

      case SIG_UNBLOCK:
        for (i = 1; i <= _NSIG; i++) {
                if (sigismember((sigset_t *)&m_in.sig_set, i))
                        sigdelset(&mp->mp_sigmask, i);
        }
        check_pending(mp);
        break;

      case SIG_SETMASK:
        sigdelset((sigset_t *) &m_in.sig_set, SIGKILL);
        mp->mp_sigmask = (sigset_t) m_in.sig_set;
        check_pending(mp);
        break;

      case SIG_INQUIRE:
        break;

      default:
        return(EINVAL);
        break;
  }
  return OK;
}

/*===========================================================================*
 *                              do_sigsuspend                                *
 *===========================================================================*/
PUBLIC int do_sigsuspend()
{
  mp->mp_sigmask2 = mp->mp_sigmask;     /* save the old mask */
  mp->mp_sigmask = (sigset_t) m_in.sig_set;
  sigdelset(&mp->mp_sigmask, SIGKILL);
  mp->mp_flags |= SIGSUSPENDED;
  check_pending(mp);
  return(SUSPEND);
}

/*===========================================================================*
 *                              do_sigreturn                                 *
 *===========================================================================*/
PUBLIC int do_sigreturn()
{
/* A user signal handler is done.  Restore context and check for
 * pending unblocked signals.
 */

  int r;

  mp->mp_sigmask = (sigset_t) m_in.sig_set;
  sigdelset(&mp->mp_sigmask, SIGKILL);

  r = sys_sigreturn(who, (struct sigmsg *) m_in.sig_context);
  check_pending(mp);
  return(r);
}

/*===========================================================================*
 *                              do_kill                                      *
 *===========================================================================*/
PUBLIC int do_kill()
{
/* Perform the kill(pid, signo) system call. */

  return check_sig(m_in.pid, m_in.sig_nr);
}

/*===========================================================================*
 *                              ksig_pending                                 *
 *===========================================================================*/
PUBLIC int ksig_pending()
{
/* Certain signals, such as segmentation violations originate in the kernel.
 * When the kernel detects such signals, it notifies the PM to take further 
 * action. The PM requests the kernel to send messages with the process
 * slot and bit map for all signaled processes. The File System, for example,
 * uses this mechanism to signal writing on broken pipes (SIGPIPE). 
 *
 * The kernel has notified the PM about pending signals. Request pending
 * signals until all signals are handled. If there are no more signals,
 * NONE is returned in the process number field.
 */ 
 int proc_nr;
 sigset_t sig_map;

 while (TRUE) {
   sys_getksig(&proc_nr, &sig_map);     /* get an arbitrary pending signal */
   if (NONE == proc_nr) {               /* stop if no more pending signals */
        break;
   } else {
        handle_sig(proc_nr, sig_map);   /* handle the received signal */
        sys_endksig(proc_nr);           /* tell kernel it's done */
   }
 } 
 return(SUSPEND);                       /* prevents sending reply */
}

/*===========================================================================*
 *                              handle_sig                                   *
 *===========================================================================*/
PRIVATE void handle_sig(proc_nr, sig_map)
int proc_nr;
sigset_t sig_map;
{
  register struct mproc *rmp;
  int i;
  pid_t proc_id, id;

  rmp = &mproc[proc_nr];
  if ((rmp->mp_flags & (IN_USE | ZOMBIE)) != IN_USE) return;
  proc_id = rmp->mp_pid;
  mp = &mproc[0];                       /* pretend signals are from PM */
  mp->mp_procgrp = rmp->mp_procgrp;     /* get process group right */

  /* Check each bit in turn to see if a signal is to be sent.  Unlike
   * kill(), the kernel may collect several unrelated signals for a
   * process and pass them to PM in one blow.  Thus loop on the bit
   * map. For SIGINT and SIGQUIT, use proc_id 0 to indicate a broadcast
   * to the recipient's process group.  For SIGKILL, use proc_id -1 to
   * indicate a systemwide broadcast.
   */
  for (i = 1; i <= _NSIG; i++) {
        if (!sigismember(&sig_map, i)) continue;
        switch (i) {
            case SIGINT:
            case SIGQUIT:
                id = 0; break;  /* broadcast to process group */
            case SIGKILL:
                id = -1; break; /* broadcast to all except INIT */
            default:
                id = proc_id;
                break;
        }
        check_sig(id, i);
  }
}

/*===========================================================================*
 *                              do_alarm                                     *
 *===========================================================================*/
PUBLIC int do_alarm()
{
/* Perform the alarm(seconds) system call. */
  return(set_alarm(who, m_in.seconds));
}

/*===========================================================================*
 *                              set_alarm                                    *
 *===========================================================================*/
PUBLIC int set_alarm(proc_nr, sec)
int proc_nr;                    /* process that wants the alarm */
int sec;                        /* how many seconds delay before the signal */
{
/* This routine is used by do_alarm() to set the alarm timer.  It is also used
 * to turn the timer off when a process exits with the timer still on.
 */
  clock_t ticks;        /* number of ticks for alarm */
  clock_t exptime;      /* needed for remaining time on previous alarm */
  clock_t uptime;       /* current system time */
  int remaining;        /* previous time left in seconds */
  int s;

  /* First determine remaining time of previous alarm, if set. */
  if (mproc[proc_nr].mp_flags & ALARM_ON) {
        if ( (s=getuptime(&uptime)) != OK) 
                panic(__FILE__,"set_alarm couldn't get uptime", s);
        exptime = *tmr_exp_time(&mproc[proc_nr].mp_timer);
        remaining = (int) ((exptime - uptime + (HZ-1))/HZ);
        if (remaining < 0) remaining = 0;       
  } else {
        remaining = 0; 
  }

  /* Tell the clock task to provide a signal message when the time comes.
   *
   * Large delays cause a lot of problems.  First, the alarm system call
   * takes an unsigned seconds count and the library has cast it to an int.
   * That probably works, but on return the library will convert "negative"
   * unsigneds to errors.  Presumably no one checks for these errors, so
   * force this call through.  Second, If unsigned and long have the same
   * size, converting from seconds to ticks can easily overflow.  Finally,
   * the kernel has similar overflow bugs adding ticks.
   *
   * Fixing this requires a lot of ugly casts to fit the wrong interface
   * types and to avoid overflow traps.  ALRM_EXP_TIME has the right type
   * (clock_t) although it is declared as long.  How can variables like
   * this be declared properly without combinatorial explosion of message
   * types?
   */
  ticks = (clock_t) (HZ * (unsigned long) (unsigned) sec);
  if ( (unsigned long) ticks / HZ != (unsigned) sec)
        ticks = LONG_MAX;       /* eternity (really TMR_NEVER) */

  if (ticks != 0) {
        pm_set_timer(&mproc[proc_nr].mp_timer, ticks, cause_sigalrm, proc_nr);
        mproc[proc_nr].mp_flags |=  ALARM_ON;
  } else if (mproc[proc_nr].mp_flags & ALARM_ON) {
        pm_cancel_timer(&mproc[proc_nr].mp_timer);
        mproc[proc_nr].mp_flags &= ~ALARM_ON;
  }
  return(remaining);
}

/*===========================================================================*
 *                              cause_sigalrm                                *
 *===========================================================================*/
PRIVATE void cause_sigalrm(tp)
struct timer *tp;
{
  int proc_nr;
  register struct mproc *rmp;

  proc_nr = tmr_arg(tp)->ta_int;        /* get process from timer */
  rmp = &mproc[proc_nr];

  if ((rmp->mp_flags & (IN_USE | ZOMBIE)) != IN_USE) return;
  if ((rmp->mp_flags & ALARM_ON) == 0) return;
  rmp->mp_flags &= ~ALARM_ON;
  check_sig(rmp->mp_pid, SIGALRM);
}

/*===========================================================================*
 *                              do_pause                                     *
 *===========================================================================*/
PUBLIC int do_pause()
{
/* Perform the pause() system call. */

  mp->mp_flags |= PAUSED;
  return(SUSPEND);
}

/*===========================================================================*
 *                              sig_proc                                     *
 *===========================================================================*/
PUBLIC void sig_proc(rmp, signo)
register struct mproc *rmp;     /* pointer to the process to be signaled */
int signo;                      /* signal to send to process (1 to _NSIG) */
{
/* Send a signal to a process.  Check to see if the signal is to be caught,
 * ignored, tranformed into a message (for system processes) or blocked.  
 *  - If the signal is to be transformed into a message, request the KERNEL to
 * send the target process a system notification with the pending signal as an 
 * argument. 
 *  - If the signal is to be caught, request the KERNEL to push a sigcontext 
 * structure and a sigframe structure onto the catcher's stack.  Also, KERNEL 
 * will reset the program counter and stack pointer, so that when the process 
 * next runs, it will be executing the signal handler. When the signal handler 
 * returns,  sigreturn(2) will be called.  Then KERNEL will restore the signal 
 * context from the sigcontext structure.
 * If there is insufficient stack space, kill the process.
 */

  vir_bytes new_sp;
  int s;
  int slot;
  int sigflags;
  struct sigmsg sm;

  slot = (int) (rmp - mproc);
  if ((rmp->mp_flags & (IN_USE | ZOMBIE)) != IN_USE) {
        printf("PM: signal %d sent to %s process %d\n",
                signo, (rmp->mp_flags & ZOMBIE) ? "zombie" : "dead", slot);
        panic(__FILE__,"", NO_NUM);
  }
  if ((rmp->mp_flags & TRACED) && signo != SIGKILL) {
        /* A traced process has special handling. */
        unpause(slot);
        stop_proc(rmp, signo);  /* a signal causes it to stop */
        return;
  }
  /* Some signals are ignored by default. */
  if (sigismember(&rmp->mp_ignore, signo)) { 
        return;
  }
  if (sigismember(&rmp->mp_sigmask, signo)) {
        /* Signal should be blocked. */
        sigaddset(&rmp->mp_sigpending, signo);
        return;
  }
  sigflags = rmp->mp_sigact[signo].sa_flags;
  if (sigismember(&rmp->mp_catch, signo)) {
        if (rmp->mp_flags & SIGSUSPENDED)
                sm.sm_mask = rmp->mp_sigmask2;
        else
                sm.sm_mask = rmp->mp_sigmask;
        sm.sm_signo = signo;
        sm.sm_sighandler = (vir_bytes) rmp->mp_sigact[signo].sa_handler;
        sm.sm_sigreturn = rmp->mp_sigreturn;
        if ((s=get_stack_ptr(slot, &new_sp)) != OK)
                panic(__FILE__,"couldn't get new stack pointer",s);
        sm.sm_stkptr = new_sp;

        /* Make room for the sigcontext and sigframe struct. */
        new_sp -= sizeof(struct sigcontext)
                                 + 3 * sizeof(char *) + 2 * sizeof(int);

        if (adjust(rmp, rmp->mp_seg[D].mem_len, new_sp) != OK)
                goto doterminate;

        rmp->mp_sigmask |= rmp->mp_sigact[signo].sa_mask;
        if (sigflags & SA_NODEFER)
                sigdelset(&rmp->mp_sigmask, signo);
        else
                sigaddset(&rmp->mp_sigmask, signo);

        if (sigflags & SA_RESETHAND) {
                sigdelset(&rmp->mp_catch, signo);
                rmp->mp_sigact[signo].sa_handler = SIG_DFL;
        }

        if (OK == (s=sys_sigsend(slot, &sm))) {

                sigdelset(&rmp->mp_sigpending, signo);
                /* If process is hanging on PAUSE, WAIT, SIGSUSPEND, tty, 
                 * pipe, etc., release it.
                 */
                unpause(slot);
                return;
        }
        panic(__FILE__, "warning, sys_sigsend failed", s);
  }
  else if (sigismember(&rmp->mp_sig2mess, signo)) {
        if (OK != (s=sys_kill(slot,signo)))
                panic(__FILE__, "warning, sys_kill failed", s);
        return;
  }

doterminate:
  /* Signal should not or cannot be caught.  Take default action. */
  if (sigismember(&ign_sset, signo)) return;

  rmp->mp_sigstatus = (char) signo;
  if (sigismember(&core_sset, signo)) {
        /* Switch to the user's FS environment and dump core. */
        tell_fs(CHDIR, slot, FALSE, 0);
        dump_core(rmp);
  }
  pm_exit(rmp, 0);              /* terminate process */
}

/*===========================================================================*
 *                              check_sig                                    *
 *===========================================================================*/
PUBLIC int check_sig(proc_id, signo)
pid_t proc_id;                  /* pid of proc to sig, or 0 or -1, or -pgrp */
int signo;                      /* signal to send to process (0 to _NSIG) */
{
/* Check to see if it is possible to send a signal.  The signal may have to be
 * sent to a group of processes.  This routine is invoked by the KILL system
 * call, and also when the kernel catches a DEL or other signal.
 */

  register struct mproc *rmp;
  int count;                    /* count # of signals sent */
  int error_code;

  if (signo < 0 || signo > _NSIG) return(EINVAL);

  /* Return EINVAL for attempts to send SIGKILL to INIT alone. */
  if (proc_id == INIT_PID && signo == SIGKILL) return(EINVAL);

  /* Search the proc table for processes to signal.  (See forkexit.c about
   * pid magic.)
   */
  count = 0;
  error_code = ESRCH;
  for (rmp = &mproc[0]; rmp < &mproc[NR_PROCS]; rmp++) {
        if (!(rmp->mp_flags & IN_USE)) continue;
        if ((rmp->mp_flags & ZOMBIE) && signo != 0) continue;

        /* Check for selection. */
        if (proc_id > 0 && proc_id != rmp->mp_pid) continue;
        if (proc_id == 0 && mp->mp_procgrp != rmp->mp_procgrp) continue;
        if (proc_id == -1 && rmp->mp_pid <= INIT_PID) continue;
        if (proc_id < -1 && rmp->mp_procgrp != -proc_id) continue;

        /* Check for permission. */
        if (mp->mp_effuid != SUPER_USER
            && mp->mp_realuid != rmp->mp_realuid
            && mp->mp_effuid != rmp->mp_realuid
            && mp->mp_realuid != rmp->mp_effuid
            && mp->mp_effuid != rmp->mp_effuid) {
                error_code = EPERM;
                continue;
        }

        count++;
        if (signo == 0) continue;

        /* 'sig_proc' will handle the disposition of the signal.  The
         * signal may be caught, blocked, ignored, or cause process
         * termination, possibly with core dump.
         */
        sig_proc(rmp, signo);

        if (proc_id > 0) break; /* only one process being signaled */
  }

  /* If the calling process has killed itself, don't reply. */
  if ((mp->mp_flags & (IN_USE | ZOMBIE)) != IN_USE) return(SUSPEND);
  return(count > 0 ? OK : error_code);
}

/*===========================================================================*
 *                              check_pending                                *
 *===========================================================================*/
PUBLIC void check_pending(rmp)
register struct mproc *rmp;
{
  /* Check to see if any pending signals have been unblocked.  The
   * first such signal found is delivered.
   *
   * If multiple pending unmasked signals are found, they will be
   * delivered sequentially.
   *
   * There are several places in this file where the signal mask is
   * changed.  At each such place, check_pending() should be called to
   * check for newly unblocked signals.
   */

  int i;

  for (i = 1; i <= _NSIG; i++) {
        if (sigismember(&rmp->mp_sigpending, i) &&
                !sigismember(&rmp->mp_sigmask, i)) {
                sigdelset(&rmp->mp_sigpending, i);
                sig_proc(rmp, i);
                break;
        }
  }
}

/*===========================================================================*
 *                              unpause                                      *
 *===========================================================================*/
PRIVATE void unpause(pro)
int pro;                        /* which process number */
{
/* A signal is to be sent to a process.  If that process is hanging on a
 * system call, the system call must be terminated with EINTR.  Possible
 * calls are PAUSE, WAIT, READ and WRITE, the latter two for pipes and ttys.
 * First check if the process is hanging on an PM call.  If not, tell FS,
 * so it can check for READs and WRITEs from pipes, ttys and the like.
 */

  register struct mproc *rmp;

  rmp = &mproc[pro];

  /* Check to see if process is hanging on a PAUSE, WAIT or SIGSUSPEND call. */
  if (rmp->mp_flags & (PAUSED | WAITING | SIGSUSPENDED)) {
        rmp->mp_flags &= ~(PAUSED | WAITING | SIGSUSPENDED);
        setreply(pro, EINTR);
        return;
  }

  /* Process is not hanging on an PM call.  Ask FS to take a look. */
  tell_fs(UNPAUSE, pro, 0, 0);
}

/*===========================================================================*
 *                              dump_core                                    *
 *===========================================================================*/
PRIVATE void dump_core(rmp)
register struct mproc *rmp;     /* whose core is to be dumped */
{
/* Make a core dump on the file "core", if possible. */

  int s, fd, seg, slot;
  vir_bytes current_sp;
  long trace_data, trace_off;

  slot = (int) (rmp - mproc);

  /* Can core file be written?  We are operating in the user's FS environment,
   * so no special permission checks are needed.
   */
  if (rmp->mp_realuid != rmp->mp_effuid) return;
  if ( (fd = open(core_name, O_WRONLY | O_CREAT | O_TRUNC | O_NONBLOCK,
                                                CORE_MODE)) < 0) return;
  rmp->mp_sigstatus |= DUMPED;

  /* Make sure the stack segment is up to date.
   * We don't want adjust() to fail unless current_sp is preposterous,
   * but it might fail due to safety checking.  Also, we don't really want 
   * the adjust() for sending a signal to fail due to safety checking.  
   * Maybe make SAFETY_BYTES a parameter.
   */
  if ((s=get_stack_ptr(slot, &current_sp)) != OK)
        panic(__FILE__,"couldn't get new stack pointer",s);
  adjust(rmp, rmp->mp_seg[D].mem_len, current_sp);

  /* Write the memory map of all segments to begin the core file. */
  if (write(fd, (char *) rmp->mp_seg, (unsigned) sizeof rmp->mp_seg)
      != (unsigned) sizeof rmp->mp_seg) {
        close(fd);
        return;
  }

  /* Write out the whole kernel process table entry to get the regs. */
  trace_off = 0;
  while (sys_trace(T_GETUSER, slot, trace_off, &trace_data) == OK) {
        if (write(fd, (char *) &trace_data, (unsigned) sizeof (long))
            != (unsigned) sizeof (long)) {
                close(fd);
                return;
        }
        trace_off += sizeof (long);
  }

  /* Loop through segments and write the segments themselves out. */
  for (seg = 0; seg < NR_LOCAL_SEGS; seg++) {
        rw_seg(1, fd, slot, seg,
                (phys_bytes) rmp->mp_seg[seg].mem_len << CLICK_SHIFT);
  }
  close(fd);
}