Skip to contentSkip to conversation

synchronous signal in the blocked signal context

DISCUSSION Siddha, Suresh B 3 participants 8 messages Aug 1, 2006

lore
Share & tools
lore
#0 lore raw

This patch (b0423a0d9cc836b2c3d796623cd19236bfedfe63)

[PATCH] Remove duplicate code in signal.c

reverts a patch introduced by Linus long time back.

http://linux.bkbits.net:8080/linux-2.6/gnupatch@3f0621871mAhWfFZzuA74eKKLvE6OQ

Was this intentional?

With the current mainline code, SIGSEGV inside a SIGSEGV handler will endup in linux handling endless recursive faults.

Just wondering if this has been discussed before and is intentional.

thanks,
suresh

#1 lore raw
Quoted 6 lines · Suresh B

This patch (b0423a0d9cc836b2c3d796623cd19236bfedfe63)

[PATCH] Remove duplicate code in signal.c

reverts a patch introduced by Linus long time back.

Good catch.

Quoted 6 lines

Was this intentional?

With the current mainline code, SIGSEGV inside a SIGSEGV handler will endup in linux handling endless recursive faults.

Just wondering if this has been discussed before and is intentional.

It certainly wasn't discussed, and I don't think it was intentional. We should _not_ just unblock a blocked signal. We should kill the process, because sending the signal is actually very very wrong.

Paul? Should I just revert, or did you have some deeper reason for it?

Linus

#2 lore raw
Quoted 17 lines · Linus Torvalds
Quoted 6 lines · Suresh B

This patch (b0423a0d9cc836b2c3d796623cd19236bfedfe63)

[PATCH] Remove duplicate code in signal.c

reverts a patch introduced by Linus long time back.

Good catch.

Quoted 5 lines

Was this intentional?

With the current mainline code, SIGSEGV inside a SIGSEGV handler will endup in linux handling endless recursive faults.

Quoted 7 lines
Quoted 1 line

Just wondering if this has been discussed before and is intentional.

It certainly wasn't discussed, and I don't think it was intentional. We should _not_ just unblock a blocked signal. We should kill the process, because sending the signal is actually very very wrong.

Paul? Should I just revert, or did you have some deeper reason for it?

I cannot claim any deep thought on this one, so please do revert it.

Next time I submit a patch to code with which I am not intimately familiar, I clearly need to carefully review the earlier patches. :-/

Thanx, Paul

#3 lore raw
Quoted 4 lines · Paul E. McKenney
Quoted 2 lines

Paul? Should I just revert, or did you have some deeper reason for it?

I cannot claim any deep thought on this one, so please do revert it.

Well, I do have to say that I like the notion of trying to have the _same_ semantics for "force_sig_info()" and "force_sig_specific()", so in that way your patch is fine - I just missed the fact that it changed it back to the old broken ones (that results in endless SIGSEGV's if the SIGSEGV happens when setting up the handler for the SIGSEGV and other "interesting" issues, where a bug can result in the user process hanging instead of just killing it outright).

However, I wonder if the _proper_ fix is to just either remove "force_sig_specific()" entirely, or just make that one match the semantics of "force_sig_info()" instead (rather than doing it the other way - change for_sig_specific() to match force_sig_info()).

force_sig_info() has only two uses, and both should be ok with the force_sig_specific() semantics, since they are for SIGSTOP and SIGKILL respectively, and those should not be blockable unless you're a kernel thread (and I don't think either of them could validly ever be used with kernel threads anyway), so doing it the other way around _should_ be ok.

Paul, Suresh, would something like this work for you instead?

Linus
----

kernel/signal.c +17 -8
index 7fe874d..bfdb568 100644
@@ -791,22 +791,31 @@ out:
 /*
  * Force a signal that the process can't ignore: if necessary
  * we unblock the signal and change any SIG_IGN to SIG_DFL.
+ *
+ * Note: If we unblock the signal, we always reset it to SIG_DFL,
+ * since we do not want to have a signal handler that was blocked
+ * be invoked when user space had explicitly blocked it.
+ *
+ * We don't want to have recursive SIGSEGV's etc, for example.
  */
-
 int
 force_sig_info(int sig, struct siginfo *info, struct task_struct *t)
 {
 	unsigned long int flags;
-	int ret;
+	int ret, blocked, ignored;
+	struct k_sigaction *action;
 
 	spin_lock_irqsave(&t->sighand->siglock, flags);
-	if (t->sighand->action[sig-1].sa.sa_handler == SIG_IGN) {
-		t->sighand->action[sig-1].sa.sa_handler = SIG_DFL;
-	}
-	if (sigismember(&t->blocked, sig)) {
-		sigdelset(&t->blocked, sig);
+	action = &t->sighand->action[sig-1];
+	ignored = action->sa.sa_handler == SIG_IGN;
+	blocked = sigismember(&t->blocked, sig);
+	if (blocked || ignored) {
+		action->sa.sa_handler = SIG_DFL;
+		if (blocked) {
+			sigdelset(&t->blocked, sig);
+			recalc_sigpending_tsk(t);
+		}
 	}
-	recalc_sigpending_tsk(t);
 	ret = specific_send_sig_info(sig, info, t);
 	spin_unlock_irqrestore(&t->sighand->siglock, flags);
 
author Acked-by
#4 lore raw
Quoted 21 lines · Linus Torvalds
Quoted 4 lines · Paul E. McKenney
Quoted 2 lines

Paul? Should I just revert, or did you have some deeper reason for it?

I cannot claim any deep thought on this one, so please do revert it.

Well, I do have to say that I like the notion of trying to have the _same_ semantics for "force_sig_info()" and "force_sig_specific()", so in that way your patch is fine - I just missed the fact that it changed it back to the old broken ones (that results in endless SIGSEGV's if the SIGSEGV happens when setting up the handler for the SIGSEGV and other "interesting" issues, where a bug can result in the user process hanging instead of just killing it outright).

However, I wonder if the _proper_ fix is to just either remove "force_sig_specific()" entirely, or just make that one match the semantics of "force_sig_info()" instead (rather than doing it the other way - change for_sig_specific() to match force_sig_info()).

Quoted 7 lines

force_sig_info() has only two uses, and both should be ok with the force_sig_specific() semantics, since they are for SIGSTOP and SIGKILL respectively, and those should not be blockable unless you're a kernel thread (and I don't think either of them could validly ever be used with kernel threads anyway), so doing it the other way around _should_ be ok.

Paul, Suresh, would something like this work for you instead?

Yes.

Quoted 47 lines

Linus
----

kernel/signal.c +17 -8
index 7fe874d..bfdb568 100644
@@ -791,22 +791,31 @@ out:
 /*
  * Force a signal that the process can't ignore: if necessary
  * we unblock the signal and change any SIG_IGN to SIG_DFL.
+ *
+ * Note: If we unblock the signal, we always reset it to SIG_DFL,
+ * since we do not want to have a signal handler that was blocked
+ * be invoked when user space had explicitly blocked it.
+ *
+ * We don't want to have recursive SIGSEGV's etc, for example.
  */
-
 int
 force_sig_info(int sig, struct siginfo *info, struct task_struct *t)
 {
 	unsigned long int flags;
-	int ret;
+	int ret, blocked, ignored;
+	struct k_sigaction *action;
 
 	spin_lock_irqsave(&t->sighand->siglock, flags);
-	if (t->sighand->action[sig-1].sa.sa_handler == SIG_IGN) {
-		t->sighand->action[sig-1].sa.sa_handler = SIG_DFL;
-	}
-	if (sigismember(&t->blocked, sig)) {
-		sigdelset(&t->blocked, sig);
+	action = &t->sighand->action[sig-1];
+	ignored = action->sa.sa_handler == SIG_IGN;
+	blocked = sigismember(&t->blocked, sig);
+	if (blocked || ignored) {
+		action->sa.sa_handler = SIG_DFL;
+		if (blocked) {
+			sigdelset(&t->blocked, sig);
+			recalc_sigpending_tsk(t);
+		}
 	}
-	recalc_sigpending_tsk(t);
 	ret = specific_send_sig_info(sig, info, t);
 	spin_unlock_irqrestore(&t->sighand->siglock, flags);
 
#5 lore raw
In reply to Linus Torvalds
Quoted 8 lines · Linus Torvalds
Quoted 4 lines · Paul E. McKenney
Quoted 2 lines

Paul? Should I just revert, or did you have some deeper reason for it?

I cannot claim any deep thought on this one, so please do revert it.

Quoted 7 lines

Well, I do have to say that I like the notion of trying to have the _same_ semantics for "force_sig_info()" and "force_sig_specific()", so in that way your patch is fine - I just missed the fact that it changed it back to the old broken ones (that results in endless SIGSEGV's if the SIGSEGV happens when setting up the handler for the SIGSEGV and other "interesting" issues, where a bug can result in the user process hanging instead of just killing it outright).

I guess I am glad I was not -totally- insane when submitting the original patch. ;-)

Quoted 4 lines

However, I wonder if the _proper_ fix is to just either remove "force_sig_specific()" entirely, or just make that one match the semantics of "force_sig_info()" instead (rather than doing it the other way - change for_sig_specific() to match force_sig_info()).

One question -- the original (2.6.14 or thereabouts) version of force_sig_info() would do the sigdelset() and recalc_sig_pending() even if the signal was not blocked, while your patch below would do sigdelset()/recalc_sig_pending() only if the signal was blocked, even if it was not ignored. Not sure this matters, but thought I should ask.

Quoted 1 line

force_sig_info() has only two uses, and both should be ok with the

s/force_sig_info/force_sig_specific/? I see >100 uses of force_sig_info().

Quoted 4 lines

force_sig_specific() semantics, since they are for SIGSTOP and SIGKILL respectively, and those should not be blockable unless you're a kernel thread (and I don't think either of them could validly ever be used with kernel threads anyway), so doing it the other way around _should_ be ok.

OK, SIGSTOP and SIGKILL cannot be ignored or blocked. So wouldn't they end up skipping the recalc_sig_pending() in the new code, where they would have ended up executing it in the 2.6.14 version of force_sig_specific()?

Assuming I am at least semi-sane, one possible way to fix shown below.

Thanx, Paul

Quoted 16 lines

Paul, Suresh, would something like this work for you instead?

Linus
----

kernel/signal.c +4 -0
index 7fe874d..bfdb568 100644
@@ -791,22 +791,31 @@ out:
 /*
  * Force a signal that the process can't ignore: if necessary
  * we unblock the signal and change any SIG_IGN to SIG_DFL.
+ *
+ * Note: If we unblock the signal, we always reset it to SIG_DFL,
+ * since we do not want to have a signal handler that was blocked
+ * be invoked when user space had explicitly blocked it.

kernel/signal.c out: · line 805

Quoted 10 lines
kernel/signal.c +3 -2
+ *
+ * We don't want to have recursive SIGSEGV's etc, for example.
  */
-
 int
 force_sig_info(int sig, struct siginfo *info, struct task_struct *t)
 {
 	unsigned long int flags;
-	int ret;
+	int ret, blocked, ignored;
	int alwaysfatal;

kernel/signal.c out: · line 810

Quoted 10 lines
kernel/signal.c +3 -5
+	struct k_sigaction *action;
 
 	spin_lock_irqsave(&t->sighand->siglock, flags);
-	if (t->sighand->action[sig-1].sa.sa_handler == SIG_IGN) {
-		t->sighand->action[sig-1].sa.sa_handler = SIG_DFL;
-	}
-	if (sigismember(&t->blocked, sig)) {
-		sigdelset(&t->blocked, sig);
+	action = &t->sighand->action[sig-1];
+	ignored = action->sa.sa_handler == SIG_IGN;
	alwaysfatal = sig == SIGKILL || sig == SIGSTOP;

kernel/signal.c out: · line 812

Quoted 2 lines
kernel/signal.c +2 -0
+	blocked = sigismember(&t->blocked, sig);
+	if (blocked || ignored) {
	if (blocked || ignored || alwaysfatal) {

kernel/signal.c out: · line 814

Quoted 2 lines
kernel/signal.c +2 -0
+		action->sa.sa_handler = SIG_DFL;
+		if (blocked) {
		if (blocked || alwaysfatal) {
Quoted 8 lines
kernel/signal.c +3 -1
+			sigdelset(&t->blocked, sig);
+			recalc_sigpending_tsk(t);
+		}
 	}
-	recalc_sigpending_tsk(t);
 	ret = specific_send_sig_info(sig, info, t);
 	spin_unlock_irqrestore(&t->sighand->siglock, flags);
 
#6 lore raw
Quoted 41 lines · Paul E. McKenney
Quoted 8 lines · Linus Torvalds
Quoted 4 lines · Paul E. McKenney
Quoted 2 lines

Paul? Should I just revert, or did you have some deeper reason for it?

I cannot claim any deep thought on this one, so please do revert it.

Quoted 7 lines

Well, I do have to say that I like the notion of trying to have the _same_ semantics for "force_sig_info()" and "force_sig_specific()", so in that way your patch is fine - I just missed the fact that it changed it back to the old broken ones (that results in endless SIGSEGV's if the SIGSEGV happens when setting up the handler for the SIGSEGV and other "interesting" issues, where a bug can result in the user process hanging instead of just killing it outright).

I guess I am glad I was not -totally- insane when submitting the original patch. ;-)

Quoted 4 lines

However, I wonder if the _proper_ fix is to just either remove "force_sig_specific()" entirely, or just make that one match the semantics of "force_sig_info()" instead (rather than doing it the other way - change for_sig_specific() to match force_sig_info()).

One question -- the original (2.6.14 or thereabouts) version of force_sig_info() would do the sigdelset() and recalc_sig_pending() even if the signal was not blocked, while your patch below would do sigdelset()/recalc_sig_pending() only if the signal was blocked, even if it was not ignored. Not sure this matters, but thought I should ask.

Quoted 1 line

force_sig_info() has only two uses, and both should be ok with the

s/force_sig_info/force_sig_specific/? I see >100 uses of force_sig_info().

Quoted 4 lines

force_sig_specific() semantics, since they are for SIGSTOP and SIGKILL respectively, and those should not be blockable unless you're a kernel thread (and I don't think either of them could validly ever be used with kernel threads anyway), so doing it the other way around _should_ be ok.

Quoted 4 lines

OK, SIGSTOP and SIGKILL cannot be ignored or blocked. So wouldn't they end up skipping the recalc_sig_pending() in the new code, where they would have ended up executing it in the 2.6.14 version of force_sig_specific()?

I don't think it matters.
signal_wake_up() in the path of specific_send_sig_info() should anyhow do that.

thanks,
suresh

#7 lore raw
Quoted 42 lines · Suresh B
Quoted 41 lines · Paul E. McKenney
Quoted 8 lines · Linus Torvalds
Quoted 4 lines · Paul E. McKenney
Quoted 2 lines

Paul? Should I just revert, or did you have some deeper reason for it?

I cannot claim any deep thought on this one, so please do revert it.

Quoted 7 lines

Well, I do have to say that I like the notion of trying to have the _same_ semantics for "force_sig_info()" and "force_sig_specific()", so in that way your patch is fine - I just missed the fact that it changed it back to the old broken ones (that results in endless SIGSEGV's if the SIGSEGV happens when setting up the handler for the SIGSEGV and other "interesting" issues, where a bug can result in the user process hanging instead of just killing it outright).

I guess I am glad I was not -totally- insane when submitting the original patch. ;-)

Quoted 4 lines

However, I wonder if the _proper_ fix is to just either remove "force_sig_specific()" entirely, or just make that one match the semantics of "force_sig_info()" instead (rather than doing it the other way - change for_sig_specific() to match force_sig_info()).

One question -- the original (2.6.14 or thereabouts) version of force_sig_info() would do the sigdelset() and recalc_sig_pending() even if the signal was not blocked, while your patch below would do sigdelset()/recalc_sig_pending() only if the signal was blocked, even if it was not ignored. Not sure this matters, but thought I should ask.

Quoted 1 line

force_sig_info() has only two uses, and both should be ok with the

s/force_sig_info/force_sig_specific/? I see >100 uses of force_sig_info().

Quoted 4 lines

force_sig_specific() semantics, since they are for SIGSTOP and SIGKILL respectively, and those should not be blockable unless you're a kernel thread (and I don't think either of them could validly ever be used with kernel threads anyway), so doing it the other way around _should_ be ok.

Quoted 8 lines
Quoted 4 lines

OK, SIGSTOP and SIGKILL cannot be ignored or blocked. So wouldn't they end up skipping the recalc_sig_pending() in the new code, where they would have ended up executing it in the 2.6.14 version of force_sig_specific()?

I don't think it matters.
signal_wake_up() in the path of specific_send_sig_info() should anyhow do that.

OK, looks plausible upon reviewing the code paths.

Thanx, Paul