Towards making KEEP_IT_REAL work again.

Fix bug that caused over-allocation of .text16 and .data16 memory areas
by a factor of 16.

src/arch/i386/include/libkir.h

 #define BASEMEM_PARAMETER_INIT BASEMEM_PARAMETER_INIT_LIBKIR
 #define BASEMEM_PARAMETER_DONE BASEMEM_PARAMETER_DONE_LIBKIR
 
+/* TEXT16_CODE: declare a fragment of code that resides in .text16 */
+#define TEXT16_CODE( asm_code_str )			\
+	".section \".text16\", \"ax\", @progbits\n\t"	\
+	".code16\n\t"					\
+	".arch i386\n\t"				\
+	asm_code_str "\n\t"				\
+	".code16gcc\n\t"				\
+	".previous\n\t"
+
 /* REAL_CODE: declare a fragment of code that executes in real mode */
 #define REAL_CODE( asm_code_str )	\
 	".code16\n\t"			\

src/arch/i386/include/librm.h

 #define BASEMEM_PARAMETER_INIT BASEMEM_PARAMETER_INIT_LIBRM
 #define BASEMEM_PARAMETER_DONE BASEMEM_PARAMETER_DONE_LIBRM
 
-/* REAL_CODE: declare a fragment of code that executes in real mode */
-#define REAL_CODE( asm_code_str )			\
-	"pushl $1f\n\t"					\
-	"call real_call\n\t"				\
-	"addl $4, %%esp\n\t"				\
+/* TEXT16_CODE: declare a fragment of code that resides in .text16 */
+#define TEXT16_CODE( asm_code_str )			\
 	".section \".text16\", \"ax\", @progbits\n\t"	\
 	".code16\n\t"					\
 	".arch i386\n\t"				\
-	"\n1:\n\t"					\
 	asm_code_str "\n\t"				\
-	"ret\n\t"					\
 	".code32\n\t"					\
 	".previous\n\t"
 
+/* REAL_CODE: declare a fragment of code that executes in real mode */
+#define REAL_CODE( asm_code_str )			\
+	"pushl $1f\n\t"					\
+	"call real_call\n\t"				\
+	"addl $4, %%esp\n\t"				\
+	TEXT16_CODE ( "\n1:\n\t"			\
+		      asm_code_str			\
+		      "\n\t"				\
+		      "ret\n\t" )
+
 #endif /* ASSEMBLY */
 
 #endif /* LIBRM_H */

src/arch/i386/kir-Makefile

 #
 CFLAGS		+= -DKEEP_IT_REAL -include kir.h
 
-# Link with _data_link_addr = 0; data symbols are relative to the data
-# segment.
-#
-LDFLAGS		+= --defsym _data_link_addr=0
-
 include Makefile
+
+LDSCRIPT	= arch/i386/scripts/i386-kir.lds

src/arch/i386/prefix/libprefix.S

 	shlw	$6, %ax
 
 	/* .data16 segment address */
-	subw	$_data16_size, %ax
+	subw	$_data16_size_pgh, %ax
 	pushw	%ax
 
 	/* .text16 segment address */
-	subw	$_text16_size, %ax
+	subw	$_text16_size_pgh, %ax
 	pushw	%ax
 
 	/* Update FBMS */
 	/* Install .text16 and .data16 */
 	call	install_basemem
 
-#ifndef KEEP_IT_REAL
+#ifdef KEEP_IT_REAL
+	/* Preserve %ds, call init_libkir, restore registers */
+	pushw	%ds
+	movw	%bx, %ds
+	movw	%ax, (init_libkir_vector+2)
+	lcall	*init_libkir_vector
+	popw	%ds
+#else
 	/* Preserve registers and interrupt status, and disable interrupts */
 	pushfw
 	pushw	%ds
 	ret
 	.size install_prealloc, . - install_prealloc
 
-#ifndef KEEP_IT_REAL
 	/* Vectors for far calls to .text16 functions */
 	.section ".data16"
+#ifdef KEEP_IT_REAL
+init_libkir_vector:
+	.word init_libkir
+	.word 0
+	.size init_libkir_vector, . - init_libkir_vector
+#else
 init_librm_vector:
 	.word init_librm
 	.word 0

src/arch/i386/scripts/i386-kir.lds

+/* -*- sh -*- */
+
+/*
+ * Linker script for i386 images
+ *
+ */
+
+OUTPUT_FORMAT ( "elf32-i386", "elf32-i386", "elf32-i386" )
+OUTPUT_ARCH ( i386 )
+ENTRY ( _entry )
+
+SECTIONS {
+
+    /* All sections in the resulting file have consecutive load
+     * addresses, but may have individual link addresses depending on
+     * the memory model being used.
+     *
+     * The linker symbols _prefix_link_addr, load_addr, and
+     * _max_align may be specified explicitly.  If not specified, they
+     * will default to:
+     *
+     *   _prefix_link_addr	= 0
+     *   _load_addr		= 0
+     *   _max_align		= 16
+     * 
+     * We guarantee alignment of virtual addresses to any alignment
+     * specified by the constituent object files (e.g. via
+     * __attribute__((aligned(x)))).  Load addresses are guaranteed
+     * only up to _max_align.  Provided that all loader and relocation
+     * code honours _max_align, this means that physical addresses are
+     * also guaranteed up to _max_align.
+     *
+     * Note that when using -DKEEP_IT_REAL, the UNDI segments are only
+     * guaranteed to be loaded on a paragraph boundary (i.e. 16-byte
+     * alignment).  Using _max_align>16 will therefore not guarantee
+     * >16-byte alignment of physical addresses when -DKEEP_IT_REAL is
+     * used (though virtual addresses will still be fully aligned).
+     *
+     */
+
+    /*
+     * The prefix
+     */
+
+    _prefix_link_addr = DEFINED ( _prefix_link_addr ) ? _prefix_link_addr : 0;
+    . = _prefix_link_addr;
+    _prefix = .;
+
+    .prefix : AT ( _prefix_load_offset + __prefix ) {
+	__prefix = .;
+	_entry = .;
+	*(.prefix)
+	*(.prefix.*)
+	_eprefix_progbits = .;
+    }
+    
+    _eprefix = .;
+
+    /*
+     * The 16-bit sections
+     */
+
+    _text16_link_addr = 0;
+    . = _text16_link_addr;
+    _text16 = .;
+
+    .text16 : AT ( _text16_load_offset + __text16 ) {
+	__text16 = .;
+	*(.text.null_trap)
+	*(.text16)
+	*(.text16.*)
+	*(.text)
+	*(.text.*)
+	_etext16_progbits = .;
+    } = 0x9090
+
+    _etext16 = .;
+
+    _data16_link_addr = 0;
+    . = _data16_link_addr;
+    _data16 = .;
+
+    .rodata16 : AT ( _data16_load_offset + __rodata16 ) {
+	__rodata16 = .;
+	*(.rodata16)
+	*(.rodata16.*)
+	*(.rodata)
+	*(.rodata.*)
+    }
+    .data16 : AT ( _data16_load_offset + __data16 ) {
+	__data16 = .;
+	*(.data16)
+	*(.data16.*)
+	*(.data)
+	*(.data.*)
+	*(SORT(.tbl.*))		/* Various tables.  See include/tables.h */
+	_edata16_progbits = .;
+    }
+    .bss16 : AT ( _data16_load_offset + __bss16 ) {
+	__bss16 = .;
+	_bss16 = .;
+	*(.bss16)
+	*(.bss16.*)
+	*(.bss)
+	*(.bss.*)
+	*(COMMON)
+	_ebss16 = .;
+    }
+    .stack16 : AT ( _data16_load_offset + __stack16 ) {
+	__stack16 = .;
+	*(.stack16)
+	*(.stack16.*)
+	*(.stack)
+	*(.stack.*)
+    }
+
+    _edata16 = .;
+
+    _end = .;
+
+    /*
+     * Dispose of the comment and note sections to make the link map
+     * easier to read
+     */
+
+    /DISCARD/ : {
+	*(.comment)
+	*(.note)
+    }
+
+    /*
+     * Load address calculations.  The slightly obscure nature of the
+     * calculations is because ALIGN(x) can only operate on the
+     * location counter.
+     */
+
+    _max_align		    = DEFINED ( _max_align ) ? _max_align : 16;
+    _load_addr		    = DEFINED ( _load_addr ) ? _load_addr : 0;
+
+    .			    = _load_addr;
+
+    .			   -= _prefix_link_addr;
+    _prefix_load_offset	    = ALIGN ( _max_align );
+    _prefix_load_addr	    = _prefix_link_addr + _prefix_load_offset;
+    _prefix_size	    = _eprefix - _prefix;
+    _prefix_progbits_size   = _eprefix_progbits - _prefix;
+    .			    = _prefix_load_addr + _prefix_progbits_size;
+
+    .			   -= _text16_link_addr;
+    _text16_load_offset	    = ALIGN ( _max_align );
+    _text16_load_addr	    = _text16_link_addr + _text16_load_offset;
+    _text16_size	    = _etext16 - _text16;
+    _text16_progbits_size   = _etext16_progbits - _text16;
+    .			    = _text16_load_addr + _text16_progbits_size;
+
+    .			   -= _data16_link_addr;
+    _data16_load_offset	    = ALIGN ( _max_align );
+    _data16_load_addr	    = _data16_link_addr + _data16_load_offset;
+    _data16_size	    = _edata16 - _data16;
+    _data16_progbits_size   = _edata16_progbits - _data16;
+    .			    = _data16_load_addr + _data16_progbits_size;
+
+    .			    = ALIGN ( _max_align );
+
+    _load_size		    = . - _load_addr;
+
+    /*
+     * Alignment checks.  ALIGN() can only operate on the location
+     * counter, so we set the location counter to each value we want
+     * to check.
+     */
+
+    . = _prefix_load_addr - _prefix_link_addr;
+    _assert = ASSERT ( ( . == ALIGN ( _max_align ) ),
+		       "_prefix is badly aligned" );
+
+    . = _text16_load_addr - _text16_link_addr;
+    _assert = ASSERT ( ( . == ALIGN ( _max_align ) ),
+		       "_text16 is badly aligned" );
+
+    . = _data16_load_addr - _data16_link_addr;
+    _assert = ASSERT ( ( . == ALIGN ( _max_align ) ),
+		       "_data16 is badly aligned" );
+
+    /*
+     * Values calculated to save code from doing it
+     */
+    _text16_size_pgh	= ( ( _text16_size + 15 ) / 16 );
+    _data16_size_pgh	= ( ( _data16_size + 15 ) / 16 );
+    _load_size_pgh	= ( ( _load_size + 15 ) / 16 );
+    _rom_size		= ( ( _load_size + 511 ) / 512 );
+}

src/arch/i386/scripts/i386.lds

      * addresses, but may have individual link addresses depending on
      * the memory model being used.
      *
-     * The linker symbols {prefix,text}_link_addr, load_addr, and
+     * The linker symbols _{prefix,textdata}_link_addr, load_addr, and
      * _max_align may be specified explicitly.  If not specified, they
      * will default to:
      *
     /*
      * Values calculated to save code from doing it
      */
-    _load_size_pgh	= ( _load_size / 16 );
+    _text16_size_pgh	= ( ( _text16_size + 15 ) / 16 );
+    _data16_size_pgh	= ( ( _data16_size + 15 ) / 16 );
+    _load_size_pgh	= ( ( _load_size + 15 ) / 16 );
     _rom_size		= ( ( _load_size + 511 ) / 512 );
 }

src/arch/i386/transitions/libkir.S

 	.section ".text16", "awx", @progbits
 	.code16
 	
+/****************************************************************************
+ * init_libkir (real-mode or 16:xx protected-mode far call)
+ *
+ * Initialise libkir ready for transitions to the kir environment
+ *
+ * Parameters:
+ *   %cs : .text16 segment
+ *   %ds : .data16 segment
+ ****************************************************************************
+ */
+	.globl	init_libkir
+init_libkir:
+	/* Record segment registers */
+	pushw	%ds
+	popw	%cs:kir_ds
+	lret
+	
 /****************************************************************************
  * ext_to_kir (real-mode or 16:xx protected-mode near call)
  *
  * %cs:0000 must point to the start of the runtime image code segment
  * on entry.
  *
- * Note that this routine can be called *without* having first set up
- * a stored kir_ds and kir_sp.  If you do this, ext_to_kir will return
- * without altering the segment registers or stack pointer.
- *
  * Parameters: none
  ****************************************************************************
  */
 	movw	%ss, %ds:ext_ss
 	movl	%esp, %ds:ext_esp
 
-	/* Load internal segment registers and stack pointer, if available */
+	/* Load internal segment registers and stack pointer */
 	movw	%ds:kir_ds, %ax
-	testw	%ax, %ax
-	jz	1f
 	movw	%ax, %ss
 	movzwl	%ds:kir_sp, %esp
 	movw	%ax, %ds
  * will also be preserved.
  *
  * Parameters:
- *   function : (16-bit) virtual address of protected-mode function to call
+ *   function : (32-bit) virtual address of C function to call
  *
  * Example usage:
- *	pushw	$pxe_api_call
+ *	pushl	$pxe_api_call
  *	lcall	$UNDI_CS, $kir_call
- *	addw	$2, %sp
+ *	addw	$4, %sp
  * to call in to the C function
  *      void pxe_api_call ( struct i386_all_regs *ix86 );
  ****************************************************************************
 
 	.globl	kir_call
 kir_call:
-
 	/* Preserve flags.  Must do this before any operation that may
 	 * affect flags.
 	 */
 	 * either a 16-bit or a 32-bit stack segment.
 	 */
 	popl	%cs:save_retaddr	/* Scratch location */
-	popw	%cs:save_function
-	subl	$6, %esp		/* Restore %esp */
+	popl	%cs:save_function
+	subl	$8, %esp		/* Restore %esp */
 	
 	/* Switch to internal stack.  Note that the external stack is
 	 * inaccessible once we're running internally (since we have
 	pushl	%cs:ext_cs_and_ss
 
 	/* Push &ix86 on stack and call function */
+	sti
 	pushl	%esp
 	data32 call *%cs:save_function
 	popl	%eax /* discard */
 		.globl kir_ds
 kir_ds:		.word 0
 		.globl kir_sp
-kir_sp:		.word 0
+kir_sp:		.word _estack
 
 /****************************************************************************
  * Temporary variables
  ****************************************************************************
  */
 save_ax:	.word 0
-save_retaddr:	.word 0
+save_retaddr:	.long 0
 save_flags:	.long 0
 save_function:	.long 0