linux/linux-5.4.31/arch/hexagon/kernel/head.S

224 lines
5.4 KiB
ArmAsm

/* SPDX-License-Identifier: GPL-2.0-only */
/*
* Early kernel startup code for Hexagon
*
* Copyright (c) 2010-2013, The Linux Foundation. All rights reserved.
*/
#include <linux/linkage.h>
#include <linux/init.h>
#include <asm/asm-offsets.h>
#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)
/*
* VMM will already have set up true vector page, MMU, etc.
* To set up initial kernel identity map, we have to pass
* the VMM a pointer to some canonical page tables. In
* this implementation, we're assuming that we've got
* them precompiled. Generate value in R24, as we'll need
* it again shortly.
*/
r24.L = #LO(swapper_pg_dir)
r24.H = #HI(swapper_pg_dir)
/*
* Symbol is kernel segment address, but we need
* the logical/physical address.
*/
r25 = pc;
r2.h = #0xffc0;
r2.l = #0x0000;
r25 = and(r2,r25); /* R25 holds PHYS_OFFSET now */
r1.h = #HI(PAGE_OFFSET);
r1.l = #LO(PAGE_OFFSET);
r24 = sub(r24,r1); /* swapper_pg_dir - PAGE_OFFSET */
r24 = add(r24,r25); /* + PHYS_OFFSET */
r0 = r24; /* aka __pa(swapper_pg_dir) */
/*
* Initialize page dir to make the virtual and physical
* addresses where the kernel was loaded be identical.
* Done in 4MB chunks.
*/
#define PTE_BITS ( __HVM_PTE_R | __HVM_PTE_W | __HVM_PTE_X \
| __HEXAGON_C_WB_L2 << 6 \
| __HVM_PDE_S_4MB)
/*
* 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 increments */
loop0(1f,r26);
1:
memw(r0 ++ #4) = r1
{ r1 = add(r1, r2); } :endloop0
/* 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. */
r31.h = #hi(__head_s_vaddr_target)
r31.l = #lo(__head_s_vaddr_target)
jumpr r31
/* Insert trippy space effects. */
__head_s_vaddr_target:
/*
* Tear down VA=PA translation now that we are running
* in kernel virtual space.
*/
r0 = #__HVM_PDE_S_INVALID
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 */
r0.h = #hi(_K_provisional_vec)
r0.l = #lo(_K_provisional_vec)
call __vmsetvec
/*
* OK, at this point we should start to be much more careful,
* we're going to enter C code and start touching memory
* in all sorts of places.
* This means:
* SGP needs to be OK
* Need to lock shared resources
* A bunch of other things that will cause
* all kinds of painful bugs
*/
/*
* Stack pointer should be pointed at the init task's
* thread stack, which should have been declared in arch/init_task.c.
* So uhhhhh...
* It's accessible via the init_thread_union, which is a union
* of a thread_info struct and a stack; of course, the top
* of the stack is not for you. The end of the stack
* is simply init_thread_union + THREAD_SIZE.
*/
{r29.H = #HI(init_thread_union); r0.H = #HI(_THREAD_SIZE); }
{r29.L = #LO(init_thread_union); r0.L = #LO(_THREAD_SIZE); }
/* initialize the register used to point to current_thread_info */
/* Fixme: THREADINFO_REG can't be R2 because of that memset thing. */
{r29 = add(r29,r0); THREADINFO_REG = r29; }
/* Hack: zero bss; */
{ r0.L = #LO(__bss_start); r1 = #0; r2.l = #LO(__bss_stop); }
{ r0.H = #HI(__bss_start); r2.h = #HI(__bss_stop); }
r2 = sub(r2,r0);
call memset;
/* Set PHYS_OFFSET; should be in R25 */
#ifdef CONFIG_HEXAGON_PHYS_OFFSET
r0.l = #LO(__phys_offset);
r0.h = #HI(__phys_offset);
memw(r0) = r25;
#endif
/* Time to make the doughnuts. */
call start_kernel
/*
* Should not reach here.
*/
1:
jump 1b
.p2align PAGE_SHIFT
ENTRY(external_cmdline_buffer)
.fill _PAGE_SIZE,1,0
.data
.p2align PAGE_SHIFT
ENTRY(empty_zero_page)
.fill _PAGE_SIZE,1,0