Hexagon: fix initial page table setup prior to jump to VA

Use the exact number of pages needed to be mapped pre-VA-jump,
then map 896MB afterwards, which the arch mem init will fix up.
Signed-off-by: Richard Kuo <rkuo@codeaurora.org>

arch/hexagon/kernel/head.S

@@ -25,6 +25,9 @@
 #include <asm/mem-layout.h>
 #include <asm/vm_mmu.h>
 #include <asm/page.h>
+#include <asm/hexagon_vm.h>
+
+#define SEGTABLE_ENTRIES #0x0e0
 
 	__INIT
 ENTRY(stext)
@@ -63,27 +66,73 @@ ENTRY(stext)
 		  | __HEXAGON_C_WB_L2 << 6			\
 		  | __HVM_PDE_S_4MB)
 
-	r1 = pc
-	r2.H = #0xffc0
-	r2.L = #0x0000
-	r1 = and(r1,r2)		/* round PC to 4MB boundary	*/
+	/*
+	 * Get number of VA=PA entries; only really needed for jump
+	 * to hyperspace; gets blown away immediately after
+	 */
+
+	{
+		r1.l = #LO(_end);
+		r2.l = #LO(stext);
+		r3 = #1;
+	}
+	{
+		r1.h = #HI(_end);
+		r2.h = #HI(stext);
+		r3 = asl(r3, #22);
+	}
+	{
+		r1 = sub(r1, r2);
+		r3 = add(r3, #-1);
+	}  /* r1 =  _end - stext  */
+	r1 = add(r1, r3);  /*  + (4M-1) */
+	r26 = lsr(r1, #22); /*  / 4M = # of entries */
+
+	r1 = r25;
+	r2.h = #0xffc0;
+	r2.l = #0x0000;		/* round back down to 4MB boundary  */
+	r1 = and(r1,r2);
 	r2 = lsr(r1, #22)	/* 4MB page number		*/
 	r2 = asl(r2, #2)	/* times sizeof(PTE) (4bytes)	*/
 	r0 = add(r0,r2)		/* r0 = address of correct PTE	*/
 	r2 = #PTE_BITS
 	r1 = add(r1,r2)		/* r1 = 4MB PTE for the first entry	*/
 	r2.h = #0x0040
-	r2.l = #0x0000		/* 4MB	*/
-	memw(r0 ++ #4) = r1
-	r1 = add(r1, r2)
+	r2.l = #0x0000		/* 4MB increments */
+	loop0(1f,r26);
+1:
 	memw(r0 ++ #4) = r1
+	{ r1 = add(r1, r2); } :endloop0
 
-	r0 = r24
+	/*  Also need to overwrite the initial 0xc0000000 entries  */
+	/*  PAGE_OFFSET >> (4MB shift - 4 bytes per entry shift)  */
+	R1.H = #HI(PAGE_OFFSET >> (22 - 2))
+	R1.L = #LO(PAGE_OFFSET >> (22 - 2))
+
+	r0 = add(r1, r24);	/* advance to 0xc0000000 entry */
+	r1 = r25;
+	r2.h = #0xffc0;
+	r2.l = #0x0000;		/* round back down to 4MB boundary  */
+	r1 = and(r1,r2);	/* for huge page */
+	r2 = #PTE_BITS
+	r1 = add(r1,r2);
+	r2.h = #0x0040
+	r2.l = #0x0000		/* 4MB increments */
+
+	loop0(1f,SEGTABLE_ENTRIES);
+1:
+	memw(r0 ++ #4) = r1;
+	{ r1 = add(r1,r2); } :endloop0
+
+	r0 = r24;
 
 	/*
 	 * The subroutine wrapper around the virtual instruction touches
 	 * no memory, so we should be able to use it even here.
+	 * Note that in this version, R1 and R2 get "clobbered"; see
+	 * vm_ops.S
 	 */
+	r1 = #VM_TRANS_TYPE_TABLE
 	call	__vmnewmap;
 
 	/*  Jump into virtual address range.  */
@@ -97,17 +146,29 @@ ENTRY(stext)
 __head_s_vaddr_target:
 	/*
 	 * Tear down VA=PA translation now that we are running
-	 * in the desgnated kernel segments.
+	 * in kernel virtual space.
 	 */
 	r0 = #__HVM_PDE_S_INVALID
-	r1 = r24
-	loop0(1f,#0x100)
+
+	r1.h = #0xffc0;
+	r1.l = #0x0000;
+	r2 = r25;		/* phys_offset */
+	r2 = and(r1,r2);
+
+	r1.l = #lo(swapper_pg_dir)
+	r1.h = #hi(swapper_pg_dir)
+	r2 = lsr(r2, #22)	/* 4MB page number		*/
+	r2 = asl(r2, #2)	/* times sizeof(PTE) (4bytes)	*/
+	r1 = add(r1,r2);
+	loop0(1f,r26)
+
 1:
 	{
 		memw(R1 ++ #4) = R0
 	}:endloop0
 
 	r0 = r24
+	r1 = #VM_TRANS_TYPE_TABLE
 	call __vmnewmap
 
 	/*  Go ahead and install the trap0 return so angel calls work  */
@@ -150,7 +211,7 @@ __head_s_vaddr_target:
 	r2 = sub(r2,r0);
 	call memset;
 
-	/*  Set PHYS_OFFSET; should still be in R25 */
+	/*  Set PHYS_OFFSET; should be in R25 */
 #ifdef CONFIG_HEXAGON_PHYS_OFFSET
 	r0.l = #LO(__phys_offset);
 	r0.h = #HI(__phys_offset);