lowlevel-z80: switch our handling of interrupts

This has implications for all platforms and breaks everything but it's worth
the pain to avoid dealing with the NMOS crawly interrupt horrors.

While breaking everything also introduce _di versions of the map functions
to allow problematic platforms to shortcut having to IRQ disable in these
cases (which are also the fast paths)

As we always save and map kernel fold those into one to optimize the IRQ
path a shade.

diff --git a/Kernel/lowlevel-z80.s b/Kernel/lowlevel-z80.s
index 7cb4346..58cc236 100644 (file)
--- a/Kernel/lowlevel-z80.s
+++ b/Kernel/lowlevel-z80.s
@@ -23,9 +23,9 @@
        .globl outnibble
 
        ; platform provided functions
-       .globl map_kernel
-       .globl map_process_always
-        .globl map_save
+       .globl map_kernel_di
+       .globl map_process_always_di
+        .globl map_save_kernel
         .globl map_restore
        .globl outchar
        .globl _inint
@@ -51,6 +51,7 @@
 
         ; imported symbols
        .globl _chksigs
+       .globl _int_disabled
         .globl _platform_monitor
         .globl _unix_syscall
         .globl outstring
@@ -81,7 +82,7 @@ deliver_signals:
        ; Pending signal
        ld a, (U_DATA__U_CURSIG)
        or a
-       ret z
+       jr z, no_pending
 
 deliver_signals_2:
        ld l, a
@@ -116,12 +117,20 @@ deliver_signals_2:
        jp (hl)         ; return to user space. This will then return via
                        ; the return path handler passed in BC
 
+no_pending:
+       ei
+       .ifne Z80_MMU_HOOKS
+       call mmu_user
+       .endif
+       ret
+
 ;
 ;      Syscall signal return path
 ;
 signal_return:
        pop hl          ; argument
-       di
+       di              ; Be careful as our internal irq state flag is
+                       ; kept wrong through this sequence.
        .ifne Z80_MMU_HOOKS
        call mmu_kernel
        .endif
@@ -132,9 +141,9 @@ signal_return:
        ;
        ld (U_DATA__U_SYSCALL_SP), sp
        ld sp, #kstack_top
-       call map_kernel
+       call map_kernel_di
        call _chksigs
-       call map_process_always
+       call map_process_always_di
        ld sp, (U_DATA__U_SYSCALL_SP)
        jr deliver_signals
 
@@ -143,6 +152,9 @@ signal_return:
 ;      Syscall processing path
 ;
 unix_syscall_entry:
+       ; We know the previous state was EI and that we won't do anything
+       ; clever until we EI again, so we can avoid the helpers on the fast
+       ; path.
         di
         ; store processor state
         ex af, af'
@@ -163,7 +175,7 @@ unix_syscall_entry:
         add hl, sp
 
        .ifne Z80_MMU_HOOKS
-       call mmu_kernel         ; must preserve HL
+       call mmu_kernel                 ; must preserve HL
        .endif
         ; save system call number
         ld a, (hl)
@@ -185,7 +197,7 @@ unix_syscall_entry:
         ld sp, #kstack_top
 
         ; map in kernel keeping common
-       call map_kernel
+       call map_kernel_di
 
         ; re-enable interrupts
         ei
@@ -201,10 +213,10 @@ unix_syscall_entry:
        ; The fork case returns with a different U_DATA mapped so the
        ; U_DATA referencing code is fine, but globals are usually not
 
-        di
+        di     ; Again we know we won't mess up calling di/ei directly
 
 
-       call map_process_always
+       call map_process_always_di
 
        xor a
        ld (U_DATA__U_INSYS), a
@@ -282,7 +294,7 @@ via_signal:
 ;
 _doexec:
         di
-        call map_process_always
+        call map_process_always_di
 
         pop bc ; return address
         pop de ; start address
@@ -350,7 +362,7 @@ nmi_handler:
        .ifne Z80_MMU_HOOKS
        call mmu_kernel
        .endif
-       call map_kernel
+       call map_kernel_di
         ld hl, #nmimsg
        jr traphl
 
@@ -411,13 +423,13 @@ mmu_irq_ret:
 
        ld a, (0)
 
-       call map_save
+       call map_save_kernel
        ;
        ;       FIXME: re-implement sanity checks and add a stack one
        ;
 
        ; We need the kernel mapped for the IRQ handling
-       call map_kernel
+       call map_kernel_di
 
        cp #0xC3
        call nz, null_pointer_trap
@@ -540,7 +552,7 @@ preemption:
        ; hence the need to reti
 
        ;
-intret2:call map_kernel
+intret2:call map_kernel_di
        ;
        ; Semantically we are doing a null syscall for pre-empt. We need
        ; to record ourselves as in a syscall so we can't be recursively
@@ -568,7 +580,7 @@ intret2:call map_kernel
        ; We have been rescheduled, remap ourself and go back to user
        ; space via signal handling
        ;
-       call map_process_always ; Get our user mapping back
+       call map_process_always_di ; Get our user mapping back
 
 
        ; We were pre-empted but have now been rescheduled
@@ -692,18 +704,35 @@ _in:
        ret
 
 ;
-;      Enable interrupts
+;      Deal with all the NMOS Z80 bugs and the buggy emulators by
+;      simply tracing our own interrupt status. It's cheaper this way
+;      but does mean any code that is using di and friends directly needs
+;      to be a lot more careful. We can also make irqflags_t 8bit and
+;      fastcall the irqrestore later on FIXME
 ;
 ___hard_ei:
+       xor a
+       ld (_int_disabled),a
        ei
        ret
 
-;
-;      Pull in the CPU specific workarounds
-;
+___hard_di:
+       ld hl,#_int_disabled
+       di
+       ld a,(hl)
+       ld (hl),#1
+       ld l,a
+       ret
 
-.ifeq CPU_NMOS_Z80
-       .include "lowlevel-z80-nmos.s"
-.else
-       .include "lowlevel-z80-cmos.s"
-.endif
+___hard_irqrestore:
+       pop de
+       pop hl
+       push hl
+       push de
+       di
+       ld a,l
+       ld (_int_disabled),a
+       or a
+       ret nz
+       ei
+       ret