From: Alan Cox <alan@linux.intel.com>
Date: Wed, 13 Dec 2017 19:14:50 +0000 (+0000)
Subject: 65c816: handle interrupt race
X-Git-Url: https://git.ndcode.org/public/gitweb.cgi?a=commitdiff_plain;h=b3e50ed2fecdcbe460c022dbff33057345af5c7b;p=FUZIX.git

65c816: handle interrupt race

Handle the 1 instruction window where we interrupt just as we do the RTL to go
back to user space. Without this additional check we see user space as the mode
and instead of executing the RTL after the CLI we end up executing something in
the user bank at that address.
---

diff --git a/Kernel/lowlevel-65c816.s b/Kernel/lowlevel-65c816.s
index 605a6743..a053f2ae 100644
--- a/Kernel/lowlevel-65c816.s
+++ b/Kernel/lowlevel-65c816.s
@@ -490,6 +490,11 @@ shoot_myself:
 ;	ZP on a 6502, fortunately on the 65C816 we can have a separate
 ;	interrupt DP
 ;
+;	Additional foulness - the 65c816 doesn't implement any way to enable
+;	interrupts one instruction on. We could avoid using rtl and stack
+;	frame for rti but the brain dead processor design makes this extra
+;	hard as the return address is differently stored!
+;
 interrupt_handler:
 	; Make sure we save all of the register bits
 	rep	#$30
@@ -543,11 +548,21 @@ join_interrupt_path:
 	txs
 
 	lda	U_DATA__U_INSYS
-	beq	ret_to_user
+	bne	ret_to_kernel
+
+	;
+	;	Peek in the frame and see if the return bank si the kernel
+	;	bank. If so its a kernel return irrespective of INSYS
+	;
+
+	lda	13,s
+	cmp 	#KERNEL_BANK
+	bne	ret_to_user
 
 	;	Kernel interrupt path may change B and D itself so we must
 	;	preserve them
 
+ret_to_kernel:
 	pld
 	plb
 
@@ -604,6 +619,7 @@ ret_to_user:
 	;	We will (one day maybe) pop back out here. It's not
 	;	guaranteed (we might be killed off)
 	;
+
 	rep	#$10
 	.a8
 	.i16
@@ -664,6 +680,11 @@ no_preempt:
 	.a8
 	.i16
 signal_exit:
+	; DEBUG FIXME
+	pha
+	lda	#1
+	sta	$FE41
+	pla
 	stz	U_DATA__U_CURSIG
 	sta	tmp1		; save signal 8bits (irq tmp1)