Skip to contentSkip to conversation

x86/xen: Return from panic notifier

PATCH Superseded Boris Ostrovsky 3 participants 7 messages Oct 1, 2019 – Oct 2, 2019

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Boris Ostrovsky

author reviewer
#0 lore raw

PATCH 1/1 x86/xen: Return from panic notifier

Currently execution of panic() continues until Xen's panic notifier (xen_panic_event()) is called at which point we make a hypercall that never returns.

This means that any notifier that is supposed to be called later as well as significant part of panic() code (such as pstore writes from kmsg_dump()) is never executed.

There is no reason for xen_panic_event() to be this last point in execution since panic()'s emergency_restart() will call into xen_emergency_restart() from where we can perform our hypercall.

---
 arch/x86/xen/enlighten.c | 17 ++++++++++++++---
 1 file changed, 14 insertions(+), 3 deletions(-)
arch/x86/xen/enlighten.c +14 -3
index 750f46ad018a..fd4f58cf51dc 100644
@@ -269,16 +269,27 @@ void xen_reboot(int reason)
 		BUG();
 }
 
+static int reboot_reason = SHUTDOWN_reboot;
 void xen_emergency_restart(void)
 {
-	xen_reboot(SHUTDOWN_reboot);
+	xen_reboot(reboot_reason);
 }
 
 static int
 xen_panic_event(struct notifier_block *this, unsigned long event, void *ptr)
 {
-	if (!kexec_crash_loaded())
-		xen_reboot(SHUTDOWN_crash);
+	if (!kexec_crash_loaded()) {
+		reboot_reason = SHUTDOWN_crash;
+
+		/*
+		 * If panic_timeout==0 then we are supposed to wait forever.
+		 * However, to preserve original dom0 behavior we have to drop
+		 * into hypervisor. (domU behavior is controlled by its
+		 * config file)
+		 */
+		if (panic_timeout == 0)
+			panic_timeout = -1;
+	}
 	return NOTIFY_DONE;
 }
 
Signature · 1 line
2.17.1
maintainer Reviewed-by on v1 1/1
#1 lore raw
Quoted 8 lines · Boris Ostrovsky

Currently execution of panic() continues until Xen's panic notifier (xen_panic_event()) is called at which point we make a hypercall that never returns.

This means that any notifier that is supposed to be called later as well as significant part of panic() code (such as pstore writes from kmsg_dump()) is never executed.

Quoted 6 lines

There is no reason for xen_panic_event() to be this last point in execution since panic()'s emergency_restart() will call into xen_emergency_restart() from where we can perform our hypercall.

Juergen

#2 lore raw
In reply to Boris Ostrovsky
Quoted 7 lines · Boris Ostrovsky

Currently execution of panic() continues until Xen's panic notifier (xen_panic_event()) is called at which point we make a hypercall that never returns.

This means that any notifier that is supposed to be called later as well as significant part of panic() code (such as pstore writes from kmsg_dump()) is never executed.

Back at the time when this was introduced into the XenoLinux tree, I think this behavior was intentional for at least DomU-s. I wonder whether you wouldn't want your change to further distinguish Dom0 and DomU behavior.

Quoted 3 lines

There is no reason for xen_panic_event() to be this last point in execution since panic()'s emergency_restart() will call into xen_emergency_restart() from where we can perform our hypercall.

Did you consider, as an alternative, to lower the notifier's priority?

Jan

author reviewer
#3 lore raw
Quoted 5 lines · Jan Beulich
Quoted 4 lines · Boris Ostrovsky

Currently execution of panic() continues until Xen's panic notifier (xen_panic_event()) is called at which point we make a hypercall that never returns.

Quoted 7 lines
Quoted 3 lines

This means that any notifier that is supposed to be called later as well as significant part of panic() code (such as pstore writes from kmsg_dump()) is never executed.

Back at the time when this was introduced into the XenoLinux tree, I think this behavior was intentional for at least DomU-s. I wonder whether you wouldn't want your change to further distinguish Dom0 and DomU behavior.

Do you remember what the reason for that was?

I think having ability to call kmsg_dump() on a panic is a useful thing to have for domUs as well. Besides, there may be other functionality added post-notifiers in panic() in the future. Or another notifier may be registered later with the same lowest priority.

Is there a downside in allowing domUs to fall through panic() all the way to emergency_restart()?

Quoted 6 lines
Quoted 4 lines

There is no reason for xen_panic_event() to be this last point in execution since panic()'s emergency_restart() will call into xen_emergency_restart() from where we can perform our hypercall.

Did you consider, as an alternative, to lower the notifier's priority?

I didn't but that wouldn't help with the original problem that James reported --- we'd still not get to kmsg_dump() call.

-boris

_______________________________________________
Xen-devel mailing list
[email protected]
https://lists.xenproject.org/mailman/listinfo/xen-devel

#4 lore raw
Quoted 9 lines · Boris Ostrovsky
Quoted 8 lines · Jan Beulich
Quoted 7 lines · Boris Ostrovsky

Currently execution of panic() continues until Xen's panic notifier (xen_panic_event()) is called at which point we make a hypercall that never returns.

This means that any notifier that is supposed to be called later as well as significant part of panic() code (such as pstore writes from kmsg_dump()) is never executed.

Quoted 6 lines
Quoted 4 lines

Back at the time when this was introduced into the XenoLinux tree, I think this behavior was intentional for at least DomU-s. I wonder whether you wouldn't want your change to further distinguish Dom0 and DomU behavior.

Do you remember what the reason for that was?

I can only guess that the thinking probably was that e.g. external dumping (by the tool stack) would be more reliable (including but not limited to this meaning less change of state from when the original crash reason was detected) than having the domain dump itself.

Quoted 7 lines

I think having ability to call kmsg_dump() on a panic is a useful thing to have for domUs as well. Besides, there may be other functionality added post-notifiers in panic() in the future. Or another notifier may be registered later with the same lowest priority.

Is there a downside in allowing domUs to fall through panic() all the way to emergency_restart()?

See above.

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

There is no reason for xen_panic_event() to be this last point in execution since panic()'s emergency_restart() will call into xen_emergency_restart() from where we can perform our hypercall.

Did you consider, as an alternative, to lower the notifier's priority?

I didn't but that wouldn't help with the original problem that James reported --- we'd still not get to kmsg_dump() call.

True. I guess more control over the behavior needs to be given to the admin, as either approach has its up- and downsides

Jan

author reviewer
#5 lore raw
Quoted 6 lines · Jan Beulich

I can only guess that the thinking probably was that e.g. external dumping (by the tool stack) would be more reliable (including but not limited to this meaning less change of state from when the original crash reason was detected) than having the domain dump itself.

We could register an external dumper (controlled by a boot option perhaps, off by default) that will call directly into hypervisor with SHUTDOWN_crash. That will guarantee that we will complete the notifier chain without relying on priorities. (Of course this still won't address a possible new feature in panic() that might be called post-dumping)

If you think it's worth doing this can be easily added.

-boris

Quoted 3 lines

True. I guess more control over the behavior needs to be given to the admin, as either approach has its up- and downsides

#6 lore raw
Quoted 9 lines · Boris Ostrovsky
Quoted 6 lines · Jan Beulich

I can only guess that the thinking probably was that e.g. external dumping (by the tool stack) would be more reliable (including but not limited to this meaning less change of state from when the original crash reason was detected) than having the domain dump itself.

Quoted 7 lines

We could register an external dumper (controlled by a boot option perhaps, off by default) that will call directly into hypervisor with SHUTDOWN_crash. That will guarantee that we will complete the notifier chain without relying on priorities. (Of course this still won't address a possible new feature in panic() that might be called post-dumping)

If you think it's worth doing this can be easily added.

Well, I think this is the better option than potentially pingponging between the two extremes, because one wants it the way it currently is and another wants it the way this patch changes things to.

Jan