kernel runs in virtual memory

This commit is contained in:
songyanguang 2024-12-19 20:05:14 +08:00
parent 8b08816b60
commit ffd2262300
3 changed files with 68 additions and 107 deletions

View File

@ -42,17 +42,16 @@ OUTPUT_FORMAT("elf64-littleriscv", "elf64-littleriscv", "elf64-littleriscv")
/* ENTRY(_start) */ /* ENTRY(_start) */
ENTRY( _boot_start ) ENTRY( _boot_start )
MEMORY { MEMORY {
vir_ddr3 (rwx) : ORIGIN = (0 - 0x80000000), LENGTH = 1024M
phy_ddr3 (rwx) : ORIGIN = 0x40200000, LENGTH = 1024M phy_ddr3 (rwx) : ORIGIN = 0x40200000, LENGTH = 1024M
vir_ddr3 (rwx) : ORIGIN = 0x0000000040800000, LENGTH = 1024M
} }
BOOT_STACK_SIZE = 0x4000; BOOT_STACK_SIZE = 0x4000;
SECTIONS SECTIONS
{ {
. = ORIGIN(phy_ddr3); . = ORIGIN(vir_ddr3);
_start = .; _start = .;
_boot_start = .; _boot_start = .;
@ -60,22 +59,22 @@ SECTIONS
_start_image_addr = .; _start_image_addr = .;
boot.o(.text) boot.o(.text)
bootmmu.o(.text .text.*)
ns16550.o(.text .text.*) ns16550.o(.text .text.*)
mmu_init.o(.text .text.*)
boot.o(.rodata .rodata.*) boot.o(.rodata .rodata.*)
bootmmu.o(.rodata .rodata.*)
ns16550.o(.rodata .rodata.*) ns16550.o(.rodata .rodata.*)
mmu_init.o(.rodata .rodata.*)
boot.o(.data .data.*) boot.o(.data .data.*)
bootmmu.o(.data .data.*)
ns16550.o(.data .data.*) ns16550.o(.data .data.*)
mmu_init.o(.data .data.*)
PROVIDE(boot_start_addr = .); PROVIDE(boot_start_addr = .);
boot.o(.bss .bss.* COMMON) boot.o(.bss .bss.* COMMON)
bootmmu.o(.bss .bss.* COMMON)
ns16550.o(.bss .bss.* COMMON) ns16550.o(.bss .bss.* COMMON)
mmu_init.o(.bss .bss.* COMMON)
. = ALIGN(0x1000); . = ALIGN(0x1000);
PROVIDE(stacks_start = .); PROVIDE(stacks_start = .);

View File

@ -35,7 +35,7 @@ Modification:
/* physical memory layout */ /* physical memory layout */
#define PHY_MEM_BASE (0x0000000040200000ULL) #define PHY_MEM_BASE (0x0000000040200000ULL)
#define PHY_USER_FREEMEM_BASE (0x0000000100000000ULL) #define PHY_USER_FREEMEM_BASE (0x0000000080000000ULL)
#define PHY_USER_FREEMEM_TOP (0x0000000200000000ULL) #define PHY_USER_FREEMEM_TOP (0x0000000200000000ULL)
#define PHY_MEM_STOP (0x0000000200000000ULL) #define PHY_MEM_STOP (0x0000000200000000ULL)
@ -78,9 +78,14 @@ Modification:
#define KERN_MEM_BASE ((0 - 0x0000002000000000ULL) + PHY_MEM_BASE) // First kernel virtual address #define KERN_MEM_BASE ((0 - 0x0000002000000000ULL) + PHY_MEM_BASE) // First kernel virtual address
#define KERN_OFFSET (KERN_MEM_BASE - PHY_MEM_BASE) #define KERN_OFFSET (KERN_MEM_BASE - PHY_MEM_BASE)
/* Leave 2GB for kernel and BPF at the end of the address space */
#define KERNEL_LINK_ADDR (0 - 0x80000000ULL)
#define KERNEL_LINK_OFFSET (KERNEL_LINK_ADDR - PHY_MEM_BASE)
/* PLIC (platform-level interrupt controller) memory layout */ /* PLIC (platform-level interrupt controller) memory layout */
#define PLIC_PHYMEM_BASE (0x0C000000ULL) #define PLIC_PHYMEM_BASE (0x0C000000ULL)
#define PLIC_MEM_SIZE (0x00400000ULL) #define PLIC_MEM_SIZE (0x00400000ULL)
#define PLIC_VIRTMEM_BASE ((0 - 0x0000003000000000ULL) + PLIC_PHYMEM_BASE)
#define V2P(a) (((uint64_t)(a)) - KERN_OFFSET) #define V2P(a) (((uint64_t)(a)) - KERN_OFFSET)
@ -88,4 +93,6 @@ Modification:
#define V2P_WO(x) ((x) - KERN_OFFSET) // same as V2P, but without casts #define V2P_WO(x) ((x) - KERN_OFFSET) // same as V2P, but without casts
#define P2V_WO(x) ((x) + KERN_OFFSET) // same as P2V, but without casts #define P2V_WO(x) ((x) + KERN_OFFSET) // same as P2V, but without casts
// clang-format on
#define V2P_LINK(a) (((uint64_t)(a)) - KERNEL_LINK_OFFSET)
#define P2V_LINK(a) ((a) + KERNEL_LINK_OFFSET)

View File

@ -33,87 +33,81 @@ Modification:
#include "pgtable.h" #include "pgtable.h"
#include "memlayout.h" #include "memlayout.h"
#include "ns16550.h" #include "ns16550.h"
#include "asm/pgtable-bits.h"
#define PFN_PD(x, prot) (((x) << _PAGE_PFN_SHIFT) | (prot))
#define _PD_PFN(x) ((x) >> _PAGE_PFN_SHIFT)
#define __ro_after_init __attribute__((section(".data..ro_after_init")))
#define __page_aligned_data __attribute__((section(".data..page_aligned"))) __attribute__((aligned(PAGE_SIZE)))
#define __page_aligned_bss __attribute__((section(".bss..page_aligned"))) __attribute__((aligned(PAGE_SIZE)))
#define __initdata __attribute__((section(".init.data")))
#define __init __attribute__((section(".init.text")))
#define __maybe_unused __attribute__((__unused__))
struct kernel_mapping kernel_map __ro_after_init;
extern char _start[]; extern char _start[];
unsigned long riscv_pfn_base __ro_after_init; struct kernel_mapping kernel_map;
pgd_t trampoline_pg_dir[PTRS_PER_PGD] __page_aligned_bss; uintptr_t trampoline_pg_dir[PTRS_PER_PGD] __attribute__((aligned(PAGE_SIZE)));
pgd_t early_pg_dir[PTRS_PER_PGD] __initdata __attribute__((aligned(PAGE_SIZE))); uintptr_t early_pg_dir[PTRS_PER_PGD] __attribute__((aligned(PAGE_SIZE)));
static pmd_t trampoline_pmd[PTRS_PER_PMD] __page_aligned_bss; static uintptr_t trampoline_pmd[PTRS_PER_PMD] __attribute__((aligned(PAGE_SIZE)));
static pmd_t early_pmd[PTRS_PER_PMD] __initdata __attribute__((aligned(PAGE_SIZE))); static uintptr_t early_pmd[PTRS_PER_PMD] __attribute__((aligned(PAGE_SIZE)));
static pmd_t early_uart_pmd[PTRS_PER_PMD] __initdata __attribute__((aligned(PAGE_SIZE))); static uintptr_t early_uart_pmd[PTRS_PER_PMD] __attribute__((aligned(PAGE_SIZE)));
static pmd_t early_pmd_free[((PHY_USER_FREEMEM_BASE - PHY_MEM_BASE) >> PGDIR_SHIFT) + 1][PTRS_PER_PMD] __initdata __attribute__((aligned(PAGE_SIZE))); static uintptr_t early_pmd_free[((PHY_USER_FREEMEM_BASE - PHY_MEM_BASE) >> PGDIR_SHIFT) + 1][PTRS_PER_PMD] __attribute__((aligned(PAGE_SIZE)));
static pmd_t early_pmd_inear_map[PTRS_PER_PMD] __initdata __attribute__((aligned(PAGE_SIZE))); static uintptr_t early_pmd_inear_map[PTRS_PER_PMD] __attribute__((aligned(PAGE_SIZE)));
static pmd_t early_plic_pmd[PTRS_PER_PMD] __initdata __attribute__((aligned(PAGE_SIZE))); static uintptr_t early_plic_pmd[PTRS_PER_PMD] __attribute__((aligned(PAGE_SIZE)));
static pmd_t *__init get_pmd_virt_early(phys_addr_t pa)
static uintptr_t *get_pmd_virt_early(phys_addr_t pa)
{ {
/* Before MMU is enabled */ /* Before MMU is enabled */
return (pmd_t *)((uintptr_t)pa); return (uintptr_t *)((uintptr_t)pa);
} }
static phys_addr_t __init alloc_pmd_early(uintptr_t va) static phys_addr_t alloc_pmd_early(uintptr_t va)
{ {
return (uintptr_t)early_pmd; return (uintptr_t)early_pmd;
} }
static void __init create_pmd_mapping_early(pmd_t *pmdp, static void create_pmd_mapping_early(uintptr_t *pmdp,
uintptr_t va, phys_addr_t pa, uintptr_t va, phys_addr_t pa,
phys_addr_t sz, pgprot_t prot) phys_addr_t sz, uintptr_t prot)
{ {
uintptr_t pmd_idx = pmd_index(va); uintptr_t pmd_idx = pmd_index(va);
if (sz == PMD_SIZE) { if (sz == PMD_SIZE) {
if (pmd_none(pmdp[pmd_idx])) if ((pmdp[pmd_idx]) == 0)
pmdp[pmd_idx] = pfn_pmd(PFN_DOWN(pa), prot); pmdp[pmd_idx] = (PFN_PD(PFN_DOWN(pa), prot));
return; return;
} }
} }
static void __init create_pgd_mapping_early(pgd_t *pgdp, static void create_pgd_mapping_early(uintptr_t *pgdp,
uintptr_t va, phys_addr_t pa, uintptr_t va, phys_addr_t pa,
phys_addr_t sz, pgprot_t prot) phys_addr_t sz, uintptr_t prot)
{ {
pmd_t *nextp; uintptr_t *nextp;
phys_addr_t next_phys; phys_addr_t next_phys;
uintptr_t pgd_idx = pgd_index(va); uintptr_t pgd_idx = pgd_index(va);
if (sz == PGDIR_SIZE) { if (sz == PGDIR_SIZE) {
if (pgd_val(pgdp[pgd_idx]) == 0) if ((pgdp[pgd_idx]) == 0)
pgdp[pgd_idx] = pfn_pgd(PFN_DOWN(pa), prot); pgdp[pgd_idx] = (PFN_PD(PFN_DOWN(pa), prot));
return; return;
} }
if (pgd_val(pgdp[pgd_idx]) == 0) { if ((pgdp[pgd_idx]) == 0) {
next_phys = alloc_pmd_early(va); next_phys = alloc_pmd_early(va);
pgdp[pgd_idx] = pfn_pgd(PFN_DOWN(next_phys), PAGE_TABLE); pgdp[pgd_idx] = (PFN_PD(PFN_DOWN(next_phys), _PAGE_TABLE));
nextp = get_pmd_virt_early(next_phys); nextp = get_pmd_virt_early(next_phys);
memset(nextp, 0, PAGE_SIZE); memset(nextp, 0, PAGE_SIZE);
} else { } else {
next_phys = PFN_PHYS(_pgd_pfn(pgdp[pgd_idx])); next_phys = PFN_PHYS(_PD_PFN((pgdp[pgd_idx])));
nextp = get_pmd_virt_early(next_phys); nextp = get_pmd_virt_early(next_phys);
} }
create_pmd_mapping_early(nextp, va, pa, sz, prot); create_pmd_mapping_early(nextp, va, pa, sz, prot);
} }
static void __init create_kernel_page_table_early(pgd_t *pgdir, bool early) static void create_kernel_page_table_early(uintptr_t *pgdir, bool early)
{ {
uintptr_t va, end_va; uintptr_t va, end_va;
@ -122,87 +116,54 @@ static void __init create_kernel_page_table_early(pgd_t *pgdir, bool early)
create_pgd_mapping_early(pgdir, va, create_pgd_mapping_early(pgdir, va,
kernel_map.phys_addr + (va - kernel_map.virt_addr), kernel_map.phys_addr + (va - kernel_map.virt_addr),
PMD_SIZE, PMD_SIZE,
PAGE_KERNEL_EXEC); (_PAGE_KERNEL | _PAGE_EXEC));
} }
} }
static void __init create_kernel_pgd_mapping_free_early(pgd_t *pgdp, static void create_kernel_pgd_mapping_free_early(uintptr_t *pgdp,
uintptr_t va, phys_addr_t pa, uintptr_t va, phys_addr_t pa,
phys_addr_t sz, pgprot_t prot) phys_addr_t sz, uintptr_t prot)
{ {
pmd_t *nextp; uintptr_t *nextp;
phys_addr_t next_phys; phys_addr_t next_phys;
uintptr_t pgd_idx = pgd_index(va); uintptr_t pgd_idx = pgd_index(va);
uintptr_t start_pgd_idx = pgd_index(kernel_map.virt_addr); uintptr_t start_pgd_idx = pgd_index(kernel_map.virt_addr);
if (pgd_val(pgdp[pgd_idx]) == 0) { if ((pgdp[pgd_idx]) == 0) {
next_phys = (uintptr_t)early_pmd_free[pgd_idx - start_pgd_idx]; next_phys = (uintptr_t)early_pmd_free[pgd_idx - start_pgd_idx];
pgdp[pgd_idx] = pfn_pgd(PFN_DOWN(next_phys), PAGE_TABLE); pgdp[pgd_idx] = (PFN_PD(PFN_DOWN(next_phys), _PAGE_TABLE));
nextp = get_pmd_virt_early(next_phys); nextp = get_pmd_virt_early(next_phys);
memset(nextp, 0, PAGE_SIZE); memset(nextp, 0, PAGE_SIZE);
} else { } else {
next_phys = PFN_PHYS(_pgd_pfn(pgdp[pgd_idx])); next_phys = PFN_PHYS(_PD_PFN((pgdp[pgd_idx])));
nextp = get_pmd_virt_early(next_phys); nextp = get_pmd_virt_early(next_phys);
} }
create_pmd_mapping_early(nextp, va, pa, sz, prot); create_pmd_mapping_early(nextp, va, pa, sz, prot);
} }
static void __init create_kernel_page_table_free_early(pgd_t *pgdir, bool early) static void create_kernel_page_table_free_early(uintptr_t *pgdir, bool early)
{ {
uintptr_t va, end_va; uintptr_t va, end_va;
end_va = kernel_map.virt_addr + (PHY_USER_FREEMEM_BASE - kernel_map.phys_addr); end_va = KERN_MEM_BASE + (PHY_USER_FREEMEM_BASE - PHY_MEM_BASE);
for (va = kernel_map.virt_addr + kernel_map.size; va < end_va; va += PMD_SIZE) { for (va = KERN_MEM_BASE + kernel_map.size; va < end_va; va += PMD_SIZE) {
create_kernel_pgd_mapping_free_early(pgdir, va, create_kernel_pgd_mapping_free_early(pgdir, va,
kernel_map.phys_addr + (va - kernel_map.virt_addr), kernel_map.phys_addr + (va - KERN_MEM_BASE),
PMD_SIZE, PMD_SIZE,
PAGE_KERNEL_EXEC); _PAGE_KERNEL);
} }
} }
static void __init create_kernel_pgd_mapping_linear_map_early(pgd_t *pgdp, static void create_plic_page_table_early(uintptr_t *pgdir, bool early)
uintptr_t va, phys_addr_t pa,
phys_addr_t sz, pgprot_t prot)
{
pmd_t *nextp;
phys_addr_t next_phys;
uintptr_t pgd_idx = pgd_index(va);
if (pgd_val(pgdp[pgd_idx]) == 0) {
next_phys = (uintptr_t)early_pmd_inear_map;
pgdp[pgd_idx] = pfn_pgd(PFN_DOWN(next_phys), PAGE_TABLE);
nextp = get_pmd_virt_early(next_phys);
memset(nextp, 0, PAGE_SIZE);
} else {
next_phys = PFN_PHYS(_pgd_pfn(pgdp[pgd_idx]));
nextp = get_pmd_virt_early(next_phys);
}
create_pmd_mapping_early(nextp, va, pa, sz, prot);
}
static void __init create_kernel_page_table_linear_map_early(pgd_t *pgdir, bool early)
{
uintptr_t va, end_va;
end_va = kernel_map.phys_addr + kernel_map.size;
for (va = kernel_map.phys_addr; va < end_va; va += PMD_SIZE) {
create_kernel_pgd_mapping_linear_map_early(pgdir, va,
kernel_map.phys_addr + (va - kernel_map.phys_addr),
PMD_SIZE,
PAGE_KERNEL_EXEC);
}
}
static void __init create_plic_page_table_early(pgd_t *pgdir, bool early)
{ {
uintptr_t va; uintptr_t va;
uintptr_t pa;
for (va = PLIC_PHYMEM_BASE; va < PLIC_PHYMEM_BASE + PLIC_MEM_SIZE; va += PMD_SIZE) { for (va = PLIC_VIRTMEM_BASE; va < PLIC_VIRTMEM_BASE + PLIC_MEM_SIZE; va += PMD_SIZE) {
create_pgd_mapping_early(pgdir, va, (uintptr_t)early_plic_pmd, PGDIR_SIZE, PAGE_TABLE); pa = va - PLIC_VIRTMEM_BASE + PLIC_PHYMEM_BASE;
create_pmd_mapping_early(early_plic_pmd, va, va, PMD_SIZE, PAGE_KERNEL); create_pgd_mapping_early(pgdir, va, (uintptr_t)early_plic_pmd, PGDIR_SIZE, _PAGE_TABLE);
create_pmd_mapping_early(early_plic_pmd, va, pa, PMD_SIZE, _PAGE_KERNEL);
} }
} }
@ -217,9 +178,9 @@ static void __init create_plic_page_table_early(pgd_t *pgdir, bool early)
* so disable compiler instrumentation when FTRACE is enabled. * so disable compiler instrumentation when FTRACE is enabled.
*/ */
void __init setup_vm_early(void) void setup_vm_early(void)
{ {
kernel_map.virt_addr = KERN_MEM_BASE; kernel_map.virt_addr = KERNEL_LINK_ADDR;
kernel_map.phys_addr = (uintptr_t)(&_start); kernel_map.phys_addr = (uintptr_t)(&_start);
kernel_map.size = (uintptr_t)(&_end) - kernel_map.phys_addr; kernel_map.size = (uintptr_t)(&_end) - kernel_map.phys_addr;
@ -227,11 +188,9 @@ void __init setup_vm_early(void)
kernel_map.va_pa_offset = PAGE_OFFSET - kernel_map.phys_addr; kernel_map.va_pa_offset = PAGE_OFFSET - kernel_map.phys_addr;
kernel_map.va_kernel_pa_offset = kernel_map.virt_addr - kernel_map.phys_addr; kernel_map.va_kernel_pa_offset = kernel_map.virt_addr - kernel_map.phys_addr;
riscv_pfn_base = PFN_DOWN(kernel_map.phys_addr);
/* Setup trampoline PGD and PMD */ /* Setup trampoline PGD and PMD */
create_pgd_mapping_early(trampoline_pg_dir, kernel_map.virt_addr, (uintptr_t)trampoline_pmd, PGDIR_SIZE, PAGE_TABLE); create_pgd_mapping_early(trampoline_pg_dir, kernel_map.virt_addr, (uintptr_t)trampoline_pmd, PGDIR_SIZE, _PAGE_TABLE);
create_pmd_mapping_early(trampoline_pmd, kernel_map.virt_addr, kernel_map.phys_addr, PMD_SIZE, PAGE_KERNEL_EXEC); create_pmd_mapping_early(trampoline_pmd, kernel_map.virt_addr, kernel_map.phys_addr, PMD_SIZE, (_PAGE_KERNEL | _PAGE_EXEC));
/* /*
* Setup early PGD covering entire kernel which will allow * Setup early PGD covering entire kernel which will allow
@ -241,16 +200,12 @@ void __init setup_vm_early(void)
create_kernel_page_table_early(early_pg_dir, true); create_kernel_page_table_early(early_pg_dir, true);
/* Setup uart PGD and PMD */ /* Setup uart PGD and PMD */
create_pgd_mapping_early(early_pg_dir, DEV_VRTMEM_BASE, (uintptr_t)early_uart_pmd, PGDIR_SIZE, PAGE_TABLE); create_pgd_mapping_early(early_pg_dir, DEV_VRTMEM_BASE, (uintptr_t)early_uart_pmd, PGDIR_SIZE, _PAGE_TABLE);
create_pmd_mapping_early(early_uart_pmd, DEV_VRTMEM_BASE, DEV_PHYMEM_BASE, PMD_SIZE, PAGE_KERNEL); create_pmd_mapping_early(early_uart_pmd, DEV_VRTMEM_BASE, DEV_PHYMEM_BASE, PMD_SIZE, _PAGE_KERNEL);
/* Setup kernel free PGD and PMD */ /* Setup kernel free PGD and PMD */
create_kernel_page_table_free_early(early_pg_dir, true); create_kernel_page_table_free_early(early_pg_dir, true);
/* Setup kernel linear map PGD and PMD */
create_kernel_page_table_linear_map_early(early_pg_dir, true);
/* Setup PLIC PGD and PMD */ /* Setup PLIC PGD and PMD */
create_plic_page_table_early(early_pg_dir, true); create_plic_page_table_early(early_pg_dir, true);
} }