Re: port-xen/58561 (panic: kernel diagnostic assertion, "x86_read_psl() == 0" failed: file, "/home/netbsd/10/src/sys/arch/x86/x86/pmap.c", line 3581)

[Manuel Bouyer <bouyer%antioche.eu.org@localhost>]

On Sat, Jan 10, 2026 at 10:01:05PM +0100, Manuel Bouyer wrote:
> On Sat, Jan 10, 2026 at 11:30:27AM -0800, Konrad Schroder wrote:
> > On 1/10/2026 4:55 AM, Manuel Bouyer wrote:
> > > Hello,
> > > can you try with the attached patch ? It won't fix the problem but
> > > should let us know if syscall() is already called with interrupts disabled,
> > > or if they're disabled later
> > 
> > Thanks!  Unfortunately that blows up almost instantly:
> > [...]
> 
> 
> Sorry; I should have checked more carefully. As there's a
> call _C_LABEL(do_pmap_load)
> early I assumed it was safe to call C functions here but do_pmap_load()
> is in fact written in assembly.
> Here's an updated patch which uses only assembly in copy.S
> Also it adds the check to all copy* functions, not only copyout.
> It boots multiuser on my test system.

And this assembly work may have allowed me to find the problem.
Basically, the CLI and STI macros are not atomic on Xen PV, and if preemption
happens at a bad time we may end up updating the upcall_mask of our previous
CPU. The attached patch should close this race condition; hopefully it's
the last one.

-- 
Manuel Bouyer <bouyer%antioche.eu.org@localhost>
     NetBSD: 26 ans d'experience feront toujours la difference
--

Index: sys/arch/amd64/include/frameasm.h
===================================================================
RCS file: /cvsroot/src/sys/arch/amd64/include/frameasm.h,v
retrieving revision 1.55
diff -u -p -u -r1.55 frameasm.h
--- sys/arch/amd64/include/frameasm.h	30 Jul 2022 14:11:00 -0000	1.55
+++ sys/arch/amd64/include/frameasm.h	10 Jan 2026 22:30:12 -0000
@@ -24,13 +24,22 @@
 #define	NOT_XEN(x)
 
 #define CLI(temp_reg) \
+ 	movq CPUVAR(CURLWP),%r ## temp_reg ;			\
+	incl L_NOPREEMPT(%r ## temp_reg);			\
  	movq CPUVAR(VCPU),%r ## temp_reg ;			\
-	movb $1,EVTCHN_UPCALL_MASK(%r ## temp_reg);
+	movb $1,EVTCHN_UPCALL_MASK(%r ## temp_reg);		\
+ 	movq CPUVAR(CURLWP),%r ## temp_reg ;			\
+	decl L_NOPREEMPT(%r ## temp_reg);			\
 
 #define STI(temp_reg) \
+ 	movq CPUVAR(CURLWP),%r ## temp_reg ;			\
+	incl L_NOPREEMPT(%r ## temp_reg);			\
  	movq CPUVAR(VCPU),%r ## temp_reg ;			\
-	movb $0,EVTCHN_UPCALL_MASK(%r ## temp_reg);
+	movb $0,EVTCHN_UPCALL_MASK(%r ## temp_reg);		\
+ 	movq CPUVAR(CURLWP),%r ## temp_reg ;			\
+	decl L_NOPREEMPT(%r ## temp_reg);			\
 
+#if 0 
 #define PUSHF(temp_reg) \
  	movq CPUVAR(VCPU),%r ## temp_reg ;			\
 	movzbl EVTCHN_UPCALL_MASK(%r ## temp_reg), %e ## temp_reg; \
@@ -39,6 +48,7 @@
 #define POPF \
 	popq %rdi; \
 	call _C_LABEL(xen_write_psl)
+#endif