X-Git-Url: https://www.chiark.greenend.org.uk/ucgi/~mdw/git/secnet/blobdiff_plain/7138d0c54cd2212439434d27cb2d6ea775c3039b..4f5e39ecfaa49376b0a5c3a4c384e91a828c1105:/process.c

diff --git a/process.c b/process.c
index b40801b..f2a2e42 100644
--- a/process.c
+++ b/process.c
@@ -5,6 +5,23 @@
 #include <sys/wait.h>
 #include "process.h"
 
+/* Advice about children from Peter:
+Better way: before the fork, make a pipe. In the child close the
++reading end. Make the writing end close-on-exec. If the dup2 or exec fails,
++write the errno value. In the parent, close the writing end. Now you can read
++from it. If you get an errno value from the pipe, the process failed and you
++know why. If you get EOF, the exec succeeded.
+
+<Senji> So, close on exec only closes if exec isn't going to return then?
+<Diziet> qu: I wouldn't bother with all that with pipes.  Remember that the
++runtime system can still make exec fail when it's `too late'.
+<Senji> Diz - I would rather have a coherant error message than 'child failed'
+<Diziet> The child, if it fails to exec, should print a message to stderr
++(giving errno and what it was trying to execute, most likely), and exit
++nonzero.
+<Diziet> It should exit calling _exit.
+*/
+
 /* Process handling - subprocesses, signals, etc. */
 
 static bool_t signal_handling=False;
@@ -39,11 +56,10 @@ static void set_default_signals(void);
    signal processing so that we can catch SIGCHLD for them and report
    their exit status using the callback function.  We block SIGCHLD
    until signal processing has begun. */
-extern void makesubproc(process_entry_fn *entry, process_callback_fn *cb,
-			void *est, void *cst, string_t desc)
+pid_t makesubproc(process_entry_fn *entry, process_callback_fn *cb,
+		 void *est, void *cst, string_t desc)
 {
     struct child *c;
-    sigset_t sigchld;
     pid_t p;
 
     c=safe_malloc(sizeof(*c),"makesubproc");
@@ -52,9 +68,7 @@ extern void makesubproc(process_entry_fn *entry, process_callback_fn *cb,
     c->cst=cst;
 
     if (!signal_handling) {
-	sigemptyset(&sigchld);
-	sigaddset(&sigchld,SIGCHLD);
-	sigprocmask(SIG_BLOCK,&sigchld,NULL);
+	fatal("makesubproc called before signal handling started");
     }
     p=fork();
     if (p==0) {
@@ -70,6 +84,7 @@ extern void makesubproc(process_entry_fn *entry, process_callback_fn *cb,
     c->finished=False;
     c->next=children;
     children=c;
+    return p;
 }
 
 static signal_notify_fn sigchld_handler;
@@ -219,11 +234,21 @@ static void set_default_signals(void)
 
 static void signal_handler(int signum)
 {
+    int saved_errno;
     uint8_t thing=0;
     sigaddset(&pending,signum);
+    /* XXX the write() may set errno, which can make the main program fail.
+       However, signal handlers aren't allowed to modify anything which
+       is not of type sig_atomic_t.  The world is broken. */
+    /* I have decided to save and restore errno anyway; on most
+       architectures on which secnet can run modifications to errno
+       will be atomic, and it seems to be the lesser of the two
+       evils. */
+    saved_errno=errno;
     write(spw,&thing,1); /* We don't care if this fails (i.e. the pipe
 			    is full) because the service routine will
 			    spot the pending signal anyway */
+    errno=saved_errno;
 }
 
 static void register_signal_handler(struct signotify *s)