More work on using sections for memory alignment. Make bcopy32 always

live in low memory so we don't have to do a bunch on-the-fly relocation,
and produce map files.

8 files changed, 62 insertions, 109 deletions

diff --git a/Makefile b/Makefile
index f0eff92e..07750a54 100644
--- a/Makefile
+++ b/Makefile
@@ -118,31 +118,32 @@ kwdhash.gen: keywords genhash.pl
 
 ldlinux.bin: ldlinux.asm kwdhash.gen version.gen
 	$(NASM) -f bin -DDATE_STR="'$(DATE)'" -DHEXDATE="$(HEXDATE)" \
-		-l ldlinux.lst -o ldlinux.bin ldlinux.asm
+		-DMAP=ldlinux.map -l ldlinux.lst -o ldlinux.bin ldlinux.asm
 
 pxelinux.bin: pxelinux.asm kwdhash.gen version.gen
 	$(NASM) -f bin -DDATE_STR="'$(DATE)'" -DHEXDATE="$(HEXDATE)" \
-		-l pxelinux.lst -o pxelinux.bin pxelinux.asm
+		-DMAP=pxelinux.map -l pxelinux.lst -o pxelinux.bin pxelinux.asm
 
 isolinux.bin: isolinux.asm kwdhash.gen version.gen checksumiso.pl
 	$(NASM) -f bin -DDATE_STR="'$(DATE)'" -DHEXDATE="$(HEXDATE)" \
-		-l isolinux.lst -o isolinux.bin isolinux.asm
+		-DMAP=isolinux.map -l isolinux.lst -o isolinux.bin isolinux.asm
 	$(PERL) checksumiso.pl isolinux.bin
 
-extlinux.bin: extlinux.asm kwdhash.gen version.gen
-	$(NASM) -f bin -DDATE_STR="'$(DATE)'" -DHEXDATE="$(HEXDATE)" \
-		-l extlinux.lst -o extlinux.bin extlinux.asm
-
-pxelinux.0: pxelinux.bin
-	cp pxelinux.bin pxelinux.0
-
 # Special verbose version of isolinux.bin
 isolinux-debug.bin: isolinux.asm kwdhash.gen version.gen checksumiso.pl
 	$(NASM) -f bin -DDATE_STR="'$(DATE)'" -DHEXDATE="$(HEXDATE)" \
 		-DDEBUG_MESSAGES \
-		-l isolinux-debug.lst -o isolinux-debug.bin isolinux.asm
+		-DMAP=isolinux-debug.map -l isolinux-debug.lst \
+		-o isolinux-debug.bin isolinux.asm
 	$(PERL) checksumiso.pl $@
 
+extlinux.bin: extlinux.asm kwdhash.gen version.gen
+	$(NASM) -f bin -DDATE_STR="'$(DATE)'" -DHEXDATE="$(HEXDATE)" \
+		-DMAP=extlinux.map -l extlinux.lst -o extlinux.bin extlinux.asm
+
+pxelinux.0: pxelinux.bin
+	cp pxelinux.bin pxelinux.0
+
 ldlinux.bss: ldlinux.bin
 	dd if=$< of=$@ bs=512 count=1
 
@@ -209,7 +210,7 @@ install-all: install install-lib
 
 local-tidy:
 	rm -f *.o *_bin.c stupid.* patch.offset
-	rm -f *.lst
+	rm -f *.lst *.map
 	rm -f $(OBSOLETE)
 
 tidy: local-tidy
diff --git a/bcopy32.inc b/bcopy32.inc
index 7b707801..5e3fda2b 100644
--- a/bcopy32.inc
+++ b/bcopy32.inc
@@ -27,11 +27,11 @@
 ; segments, but this stuff is painful enough as it is without having to rely
 ; on everything happening "as it ought to."
 ;
-; IMPORTANT: This code must be capable of operating when copied to the
-; trackbuf area (1000h).  The routine bcopy_over_self handles this mode
-; of operation, including any necessary adjustments.
+; NOTE: this code is relocated into low memory, just after the .earlybss
+; segment, in order to support to "bcopy over self" operation.
 ;
-		section .text
+
+		section .bcopy32
 		align 8
 __bcopy_start:
 
@@ -39,7 +39,7 @@ __bcopy_start:
 		; contiguous with the rest of the bcopy stuff
 
 bcopy_gdt:	dw bcopy_gdt_size-1	; Null descriptor - contains GDT
-.adj1:		dd bcopy_gdt		; pointer for LGDT instruction
+		dd bcopy_gdt		; pointer for LGDT instruction
 		dw 0
 		dd 0000ffffh		; Code segment, use16, readable,
 		dd 00009b00h		; present, dpl 0, cover 64K
@@ -81,11 +81,11 @@ bcopy:		push eax
 		cli
 		call enable_a20
 
-.adj2:		o32 lgdt [cs:bcopy_gdt]
+		o32 lgdt [cs:bcopy_gdt]
 		mov eax,cr0
 		or al,1
 		mov cr0,eax		; Enter protected mode
-.adj3a:		jmp 08h:.in_pm
+		jmp 08h:.in_pm
 
 .in_pm:		mov ax,10h		; Data segment selector
 		mov es,ax
@@ -112,7 +112,7 @@ bcopy:		push eax
 		mov eax,cr0
 		and al,~1
 		mov cr0,eax		; Disable protected mode
-.adj3b:		jmp 0:.in_rm
+		jmp 0:.in_rm
 
 .in_rm:		; Back in real mode
 		lss sp,[cs:SavedSSSP]
@@ -145,11 +145,6 @@ bcopy:		push eax
 %define A20_KBC		3		; A20 through KBC
 %define A20_FAST	4		; A20 through port 92h
 
-		align 2
-A20List		dw a20_dunno, a20_none, a20_bios, a20_kbc, a20_fast
-A20DList	dw a20d_dunno, a20d_none, a20d_bios, a20d_kbc, a20d_fast
-a20_adjust_cnt	equ ($-A20List)/2
-
 slow_out:	out dx, al		; Fall through
 
 _io_delay:	out IO_DELAY_PORT,al
@@ -173,7 +168,7 @@ try_enable_a20:
 ;
 		mov bp,[cs:A20Type]
 		add bp,bp			; Convert to word offset
-.adj4:		jmp word [cs:bp+A20List]
+		jmp word [cs:bp+A20List]
 
 ;
 ; First, see if we are on a system with no A20 gate
@@ -299,7 +294,7 @@ disable_a20:
 
 		mov bp,[cs:A20Type]
 		add bp,bp			; Convert to word offset
-.adj5:		jmp word [cs:bp+A20DList]
+		jmp word [cs:bp+A20DList]
 
 a20d_bios:
 		mov ax,2400h
@@ -396,52 +391,7 @@ try_wbinvd:
 ;	ESI, EDI, ECX	- same as bcopy
 ;	On stack	- initial state (fd, ad, ds, es, fs, gs)
 ;
-ADJUST		equ (__bcopy_start - $$) + TEXT_START - BSS_START
-
-		align 2
-adjlist		dw bcopy_gdt.adj1 - ADJUST
-		dw bcopy.adj2 + 5 - ADJUST
-		dw bcopy.adj3a + 1 - ADJUST
-		dw bcopy.adj3b + 1 - ADJUST
-		dw try_enable_a20.adj4 + 3 - ADJUST
-		dw disable_a20.adj5 + 3 - ADJUST
-adjlist_cnt	equ ($-adjlist)/2
-
 bcopy_over_self:
-		push esi	
-		push edi
-		push ecx
-
-		xor bx,bx
-		mov es,bx
-		mov ds,bx
-
-		mov si,__bcopy_start
-		mov di,trackbuf		; == BSS_START
-		mov cx,(__bcopy_end - __bcopy_start + 3) >> 2
-		rep movsd
-
-		mov si,A20List - ADJUST
-		mov cx,a20_adjust_cnt
-.adjust1:
-		sub word [si], ADJUST
-		inc si
-		inc si
-		loop .adjust1
-
-		mov si, adjlist
-		mov cx, adjlist_cnt
-.adjust2:
-		lodsw
-		xchg di,ax
-		sub word [di], ADJUST
-		loop .adjust2
-
-		jmp .next-ADJUST
-.next:
-		pop ecx
-		pop edi
-		pop esi
 		call bcopy
 		
 		pop gs
@@ -451,9 +401,17 @@ bcopy_over_self:
 		popad
 		popfd
 		jmp 0:7c00h
-__bcopy_end:
 
-		section .earlybss
-A20Test		resw 1			; Counter for testing status of A20
-A20Type		resw 1			; A20 type
-A20Tries	resb 1			; Times until giving up on A20
+		align 2
+A20List		dw a20_dunno, a20_none, a20_bios, a20_kbc, a20_fast
+A20DList	dw a20d_dunno, a20d_none, a20d_bios, a20d_kbc, a20d_fast
+a20_adjust_cnt	equ ($-A20List)/2
+
+A20Type		dw A20_NONE		; A20 type
+A20Test		dw 0			; Counter for testing status of A20
+A20Tries	db 0			; Times until giving up on A20
+
+		; Total size of .bcopy32 section
+		alignb 4, db 0		; Even number of dwords
+__bcopy_size	equ $-__bcopy_start
+
diff --git a/cpuinit.inc b/cpuinit.inc
index 866c5f3a..8ac20f51 100644
--- a/cpuinit.inc
+++ b/cpuinit.inc
@@ -41,6 +41,14 @@ enough_ram:
 skip_checks:
 
 ;
+; Initialize the bcopy32 code in low memory
+;
+		mov si,section..bcopy32.start
+		mov di,__bcopy_start
+		mov cx,__bcopy_size >> 2
+		rep movsd
+
+;
 ; Check if we're 386 (as opposed to 486+); if so we need to blank out
 ; the WBINVD instruction
 ;
@@ -65,10 +73,4 @@ skip_checks:
 		mov byte [try_wbinvd],0c3h		; Near RET
 is_486:
 %endif	; DO_WBINVD
-;
-; Mark A20 type as unknown.  This particular word needs to be in BSS,
-; so it needs to be initialized.
-;
-		mov word [A20Type],0
-
 
diff --git a/extlinux.asm b/extlinux.asm
index fcd3fc30..dac3393f 100644
--- a/extlinux.asm
+++ b/extlinux.asm
@@ -22,6 +22,7 @@
 %include "kernel.inc"
 %include "bios.inc"
 %include "tracers.inc"
+%include "layout.inc"
 
 %include "ext2_fs.inc"
 
@@ -107,14 +108,13 @@ file_mode	resw 1
 ;
 ; Memory below this point is reserved for the BIOS and the MBR
 ;
-BSS_START	equ 0800h
-		section .earlybss nobits start=BSS_START
+		section .earlybss
 trackbufsize	equ 8192
 trackbuf	resb trackbufsize	; Track buffer goes here
 getcbuf		resb trackbufsize
 		; ends at 4800h
 
-		section .bss nobits align=256 follows=.earlybss
+		section .bss
 SuperBlock	resb 1024		; ext2 superblock
 SuperInfo	resq 16			; DOS superblock expanded
 ClustSize	resd 1			; Bytes/cluster ("block")
@@ -142,8 +142,6 @@ xbs_vgatmpbuf	equ 2*trackbufsize
 
 
 		section .text
-TEXT_START	equ 7C00h
-                org TEXT_START
 ;
 ; Some of the things that have to be saved very early are saved
 ; "close" to the initial stack pointer offset, in order to
diff --git a/font.inc b/font.inc
index 9600b30f..d3dbf456 100644
--- a/font.inc
+++ b/font.inc
@@ -120,7 +120,7 @@ vidrows_ok:	mov [VidRows],al
 		
 ; VGA font buffer at the end of memory (so loading a font works even
 ; in graphics mode, but don't put too much pressure on the .bss)
-		section vgafont nobits start=0E000h
+		section .latebss
 vgafontbuf	resb 8192
 
 		section .data
diff --git a/isolinux.asm b/isolinux.asm
index 3ba82972..4a3af1de 100644
--- a/isolinux.asm
+++ b/isolinux.asm
@@ -25,6 +25,7 @@
 %include "kernel.inc"
 %include "bios.inc"
 %include "tracers.inc"
+%include "layout.inc"
 
 ;
 ; Some semi-configurable constants... change on your own risk.
@@ -103,14 +104,13 @@ dir_clust	resd 1			; Length in clusters
 ;
 ; Memory below this point is reserved for the BIOS and the MBR
 ;
-BSS_START	equ 0800h
-		section .earlybss nobits start=BSS_START
+		section .earlybss
 trackbufsize	equ 8192
 trackbuf	resb trackbufsize	; Track buffer goes here
 getcbuf		resb trackbufsize
 ;		ends at 4800h
 
-		section .bss nobits align=256 follows=.earlybss
+		section .bss
 		alignb 4
 ISOFileName	resb 64			; ISO filename canonicalization buffer
 ISOFileNameEnd	equ $
@@ -141,8 +141,6 @@ xbs_vgabuf	equ trackbufsize
 xbs_vgatmpbuf	equ 2*trackbufsize
 
 		section .text
-TEXT_START	equ 7C00h
-                org TEXT_START
 ;;
 ;; Primary entry point.  Because BIOSes are buggy, we only load the first
 ;; CD-ROM sector (2K) of the file, so the number one priority is actually
diff --git a/ldlinux.asm b/ldlinux.asm
index 2a5dfcc6..e528adb4 100644
--- a/ldlinux.asm
+++ b/ldlinux.asm
@@ -30,7 +30,7 @@
 %include "kernel.inc"
 %include "bios.inc"
 %include "tracers.inc"
-
+%include "layout.inc"
 ;
 ; Some semi-configurable constants... change on your own risk.
 ;
@@ -109,14 +109,13 @@ file_left	resd 1			; Number of sectors left
 ;
 ; Memory below this point is reserved for the BIOS and the MBR
 ;
-BSS_START	equ 0800h
-		section .earlybss nobits start=BSS_START
+		section .earlybss
 trackbufsize	equ 8192
 trackbuf	resb trackbufsize	; Track buffer goes here
 getcbuf		resb trackbufsize
 		; ends at 4800h
 
-		section .bss nobits align=256 follows=.earlybss
+		section .bss
 		alignb 8
 
 		; Expanded superblock
@@ -152,8 +151,6 @@ xbs_vgatmpbuf	equ 2*trackbufsize
 
 
 		section .text
-TEXT_START	equ 7C00h
-                org TEXT_START
 ;
 ; Some of the things that have to be saved very early are saved
 ; "close" to the initial stack pointer offset, in order to
diff --git a/pxelinux.asm b/pxelinux.asm
index 2b7819bd..f99478e4 100644
--- a/pxelinux.asm
+++ b/pxelinux.asm
@@ -25,6 +25,7 @@
 %include "bios.inc"
 %include "tracers.inc"
 %include "pxe.inc"
+%include "layout.inc"
 
 ;
 ; Some semi-configurable constants... change on your own risk.
@@ -175,8 +176,7 @@ tftp_pktbuf	resw 1			; Packet buffer offset
 ;
 ; Memory below this point is reserved for the BIOS and the MBR
 ;
-BSS_START	equ 0800h
- 		section .earlybss nobits start=BSS_START
+ 		section .earlybss
 trackbufsize	equ 8192
 trackbuf	resb trackbufsize	; Track buffer goes here
 getcbuf		resb trackbufsize
@@ -203,7 +203,7 @@ IPOption	resb 80			; ip= option buffer
 ; writing a received ARP packet into low memory.
 RBFG_brainfuck	resb 0E00h
 
-		section .bss nobits align=256 follows=.earlybss
+		section .bss
 		alignb 4
 InitStack	resd 1			; Pointer to reset stack
 RebootTime	resd 1			; Reboot timeout, if set by option
@@ -255,8 +255,6 @@ xbs_vgatmpbuf	equ 2*trackbufsize
 		; PXELINUX needs more BSS than the other derivatives;
 		; therefore we relocate it from 7C00h on startup
 		;
-TEXT_START	equ 9000h
-                org TEXT_START
 StackBuf	equ $-44		; Base of stack if we use our own
 
 ;
@@ -271,21 +269,22 @@ _start:
 		push fs
 		push gs
 
+		xor ax,ax
+		mov ds,ax
+		mov es,ax	
+
 		mov si,ldlinux_end-(TEXT_START-7C00h)-4
 		mov di,ldlinux_end-4
 		mov cx,ldlinux_end-TEXT_START
 		shr cx,2
 		std			; Overlapping areas, copy backwards
-		rep movsb
+		rep movsd
 
 		jmp 0:_start1		; Canonicalize address
 _start1:
 		mov bp,sp
 		les bx,[bp+48]		; ES:BX -> !PXE or PXENV+ structure
 
-		mov ax,cs
-		mov ds,ax
-
 		; That is all pushed onto the PXE stack.  Save the pointer
 		; to it and switch to an internal stack.
 		mov [InitStack],sp