kernel: run with interrupts enabled when possible

[msft-jlange]

On platforms that support SVSM interrupt delivery correctly (all platforms except SNP without Restricted Injection), the kernel should generally run with interrupts enabled so that interrupts directed at the SVSM can be handled as quickly as possible.

[joergroedel]

The changes break test-in-svsm in my setup, I need to investigate before merging.

[stefano-garzarella]

The changes break test-in-svsm in my setup, I need to investigate before merging.

Yeah, also on my env:

[SVSM] Launching request-processing task on CPU 0
[SVSM] running 141 tests
[SVSM] test cpu::irq_state::tests::irq_enable_disable ...
[SVSM] ERROR: Panic: CPU[0] panicked at kernel/src/cpu/irq_state.rs:228:9:
assertion failed: irqs_disabled()
[SVSM] ---BACKTRACE---:
[SVSM]   [0xffffff8000112845]
[SVSM]   [0xffffff80000cecb3]
[SVSM]   [0xffffff8000014201]
[SVSM]   [0xffffff8000025021]
[SVSM]   [0xffffff80000ae256]
[SVSM]   [0xffffff800007a795]
[SVSM]   [0xffffff80000377e0]
[SVSM]   [0xffffff800006b52b]
[SVSM]   [0xffffff800006b540]
[SVSM] ---END---

Maybe with commit 78b0a92 we need to adjust tests in kernel/src/cpu/irq_state.rs since we don't start with irq disabled anymore. (I didn't follow the details, just guessing)

[msft-jlange]

The changes break test-in-svsm in my setup, I need to investigate before merging.

I've modified the tests to understand the interrupt configuration, so hopefully everything will work now.

[stefano-garzarella]

The changes break test-in-svsm in my setup, I need to investigate before merging.

I've modified the tests to understand the interrupt configuration, so hopefully everything will work now.

Some other tests failing in kernel/src/locking/rwlock.rs :

[SVSM] test locking::rwlock::tests::rw_lock_irq_safe ...
[SVSM] ERROR: Panic: CPU[0] panicked at kernel/src/locking/rwlock.rs:368:9:
assertion failed: irqs_disabled()
[SVSM] ---BACKTRACE---:
[SVSM]   [0xffffff8000112855]
[SVSM]   [0xffffff8000010b46]
[SVSM]   [0xffffff80000210a1]
[SVSM]   [0xffffff8000024871]
[SVSM]   [0xffffff80000ae256]
[SVSM]   [0xffffff800007a795]
[SVSM]   [0xffffff80000377e0]
[SVSM]   [0xffffff800006b52b]
[SVSM]   [0xffffff800006b540]
[SVSM] ---END---
qemu-system-x86_64: terminating on signal 2

But applying the same changes you did also there, everything is fine:

diff --git a/kernel/src/locking/rwlock.rs b/kernel/src/locking/rwlock.rs
index 9d81b7d..f48fa55 100644
--- a/kernel/src/locking/rwlock.rs
+++ b/kernel/src/locking/rwlock.rs
@@ -330,11 +330,12 @@ mod tests {
     #[cfg_attr(not(test_in_svsm), ignore = "Can only be run inside guest")]
     fn rw_lock_irq_unsafe() {
         use crate::cpu::irq_state::{raw_irqs_disable, raw_irqs_enable};
-        use crate::cpu::{irqs_disabled, irqs_enabled};
+        use crate::cpu::{irqs_enabled};
         use crate::locking::*;
 
-        assert!(irqs_disabled());
         unsafe {
+            let was_enabled = irqs_enabled();
+
             raw_irqs_enable();
             let lock = RWLock::new(0);
 
@@ -355,6 +356,10 @@ mod tests {
             // IRQs must still be enabled
             assert!(irqs_enabled());
             raw_irqs_disable();
+
+            if !was_enabled {
+                raw_irqs_disable();
+            }
         }
     }
 
@@ -365,8 +370,9 @@ mod tests {
         use crate::cpu::{irqs_disabled, irqs_enabled};
         use crate::locking::*;
 
-        assert!(irqs_disabled());
         unsafe {
+            let was_enabled = irqs_enabled();
+
             raw_irqs_enable();
             let lock = RWLockIrqSafe::new(0);
 
@@ -389,6 +395,10 @@ mod tests {
             // IRQs must still be enabled
             assert!(irqs_enabled());
             raw_irqs_disable();
+
+            if !was_enabled {
+                raw_irqs_disable();
+            }
         }
     }
 }

BTW, in order not to break the bisection, I think it's better to change the tests before commit 78b0a92 or along with it.

[msft-jlange]

Some other tests failing in kernel/src/locking/rwlock.rs :

Thanks for identifying the problem!

BTW, in order not to break the bisection, I think it's better to change the tests before commit 78b0a92 or along with it.

Fixed.

[stefano-garzarella]

Another issue :-( I guess because we merged new stuff in main:

[SVSM] ERROR: Panic: CPU[0] panicked at kernel/src/locking/spinlock.rs:245:9:
assertion failed: irqs_disabled()

There are some tests in kernel/src/locking/spinlock.rs with the same assumption (irq disabled) that we need to fix.

nit: maybe better to update the title of commit 05082b0 since we are fixing tests not only in cpu/irq_state

[msft-jlange]

There are some tests in kernel/src/locking/spinlock.rs with the same assumption (irq disabled) that we need to fix.

Thanks for checking! Hopefully I've got them all now.

nit: maybe better to update the title of commit 05082b0 since we are fixing tests not only in cpu/irq_state

There's no obviously good name because there isn't a clear path here, so hopefully the new title of the commit is good enough.

[stefano-garzarella]

There are some tests in kernel/src/locking/spinlock.rs with the same assumption (irq disabled) that we need to fix.

Thanks for checking! Hopefully I've got them all now.

Yeah, no failures running make test-in-svsm on my env.

nit: maybe better to update the title of commit 05082b0 since we are fixing tests not only in cpu/irq_state

There's no obviously good name because there isn't a clear path here, so hopefully the new title of the commit is good enough.

Yep, I think it's fine ;-)

Thanks!