add usermode jump into init

1 files changed, 127 insertions, 15 deletions

diff --git a/arch/x86_64/sys/init.c b/arch/x86_64/sys/init.c
index 06fbe48..2bd86d5 100644
--- a/arch/x86_64/sys/init.c
+++ b/arch/x86_64/sys/init.c
@@ -1,42 +1,75 @@
 #include "init.h"
+#include "arch/x86_64/hal/msr.h"
 #include "config.h"
 #include "jove.h"
 #include "hal/obj.h"
 #include "lib/string.h"
 #include "lib/panic.h"
+#include "lib/printf.h"
+#include "arch/x86_64/lib/elf.h"
 #include "arch/x86_64/hal/paging.h"
 #include "boot/cmdline.h"
 #include "boot/initrd.h"
 
-#define INIT_PROGRAM_HUGEPAGES_MAX ((INIT_PROGRAM_PAGES_MAX + 511) / 512)
-#define INIT_STACK_HUGEPAGES_MAX ((INIT_STACK_PAGES_MAX + 511) / 512)
-
 PAGEALIGN static page_t s_init_program_pml3[512];
 PAGEALIGN static page_t s_init_program_pml2[512];
-PAGEALIGN static page_t s_init_program_pml1[INIT_PROGRAM_HUGEPAGES_MAX][512];
+PAGEALIGN static page_t s_init_program_pml1[512];
+PAGEALIGN static uint8_t s_init_program_pages[INIT_PROGRAM_PAGES_MAX][0x1000];
 
 PAGEALIGN static page_t s_init_stack_pml3[512];
 PAGEALIGN static page_t s_init_stack_pml2[512];
-PAGEALIGN static page_t s_init_stack_pml1[INIT_STACK_HUGEPAGES_MAX][512];
+PAGEALIGN static page_t s_init_stack_pml1[512];
+PAGEALIGN static uint8_t s_init_stack_pages[INIT_STACK_PAGES][0x1000];
+
+static int
+elf_is_valid(struct elf_header *ehdr)
+{
+    if(ehdr->e_ident[0] != EI_MAG0 ||
+       ehdr->e_ident[1] != EI_MAG1 ||
+       ehdr->e_ident[2] != EI_MAG2 ||
+       ehdr->e_ident[3] != EI_MAG3) {
+        kprintf("ELF: Magic does not match!\n");
+        return 0;
+    }
+    if(ehdr->e_ident[4] != EI_CLASS_64) {
+        kprintf("ELF: Class is not 64-bit!\n");
+        return 0;
+    }
+    if(ehdr->e_ident[5] != EI_DATA_LE) {
+        kprintf("ELF: Data is not little-endian!\n");
+        return 0;
+    }
+    if(ehdr->e_type != ET_EXEC) {
+        kprintf("ELF: Program is not standard executable type!\n");
+        return 0;
+    }
+    if(ehdr->e_machine != EM_x86_64) {
+        kprintf("ELF: Executable is not x86_64!\n");
+        return 0;
+    }
+    return 1;
+}
 
 void
 init_setup(void)
 {
     memset(s_init_program_pml3, 0, 0x1000);
     memset(s_init_program_pml2, 0, 0x1000);
-    memset(s_init_program_pml1, 0, 0x1000 * INIT_PROGRAM_HUGEPAGES_MAX);
+    memset(s_init_program_pml1, 0, 0x1000);
 
     memset(s_init_stack_pml3, 0, 0x1000);
     memset(s_init_stack_pml2, 0, 0x1000);
-    memset(s_init_stack_pml1, 0, 0x1000 * INIT_STACK_HUGEPAGES_MAX);
+    memset(s_init_stack_pml1, 0, 0x1000);
 
-    uintptr_t init_program_pml3_pptr = KPTR_TOPHYS(s_init_program_pml3);
-    uintptr_t init_program_pml2_pptr = KPTR_TOPHYS(s_init_program_pml2);
-    uintptr_t init_program_pml1_pptr = KPTR_TOPHYS(s_init_program_pml1);
+    uintptr_t init_program_pml3_pptr = KPTR_TOPHYS(&s_init_program_pml3);
+    uintptr_t init_program_pml2_pptr = KPTR_TOPHYS(&s_init_program_pml2);
+    uintptr_t init_program_pml1_pptr = KPTR_TOPHYS(&s_init_program_pml1);
+    uintptr_t init_program_pml0_pptr = KPTR_TOPHYS(&s_init_program_pages);
 
-    uintptr_t init_stack_pml3_pptr = KPTR_TOPHYS(s_init_stack_pml3);
-    uintptr_t init_stack_pml2_pptr = KPTR_TOPHYS(s_init_stack_pml2);
-    uintptr_t init_stack_pml1_pptr = KPTR_TOPHYS(s_init_stack_pml1);
+    uintptr_t init_stack_pml3_pptr = KPTR_TOPHYS(&s_init_stack_pml3);
+    uintptr_t init_stack_pml2_pptr = KPTR_TOPHYS(&s_init_stack_pml2);
+    uintptr_t init_stack_pml1_pptr = KPTR_TOPHYS(&s_init_stack_pml1);
+    uintptr_t init_stack_pml0_pptr = KPTR_TOPHYS(&s_init_stack_pages);
 
     //Get the init file data from the initrd.
     const char *init_path = cmdline_getarg("init");
@@ -45,9 +78,88 @@ init_setup(void)
     }
 
     jove_owned_obj_t *initrd_obj = obj_take(objdir_getlocal(_rootdir_obj, ROOTDIR_INITRD_I));
-    jove_owned_obj_t *initfile_obj = initrd_seek(initrd_obj, init_path);
-    if(initfile_obj == NULL) {
+    jove_owned_obj_t *init_obj = initrd_seek(initrd_obj, init_path);
+    if(init_obj == NULL) {
         kpanic("init: Cannot find file '%s' in initrd.\n", init_path);
     }
     obj_release(initrd_obj);
+
+    void *init_data = initrd_file_data(init_obj);
+    obj_release(init_obj);
+
+    struct elf_header *init_ehdr = init_data;
+    if(init_ehdr == NULL) {
+        kpanic("init: Cannot get file data for file '%s': %i\n", init_path, errno);
+    }
+
+    if(!elf_is_valid(init_ehdr)) {
+        kpanic("init: Either passed file '%s' is not an ELF file or is not loadable.\n", init_path);
+    }
+
+    jove_owned_obj_t *init_pagedir = obj_take(objdir_getlocal(_rootdir_obj, ROOTDIR_PAGEDIR_I));
+    page_t *pml4 = obj_deref(init_pagedir);
+
+    //Load the init ELF
+    for(size_t i = 0; i < init_ehdr->e_shnum; i++) {
+        struct elf_shdr *shdr = init_data + init_ehdr->e_shoff + i * init_ehdr->e_shentsize;
+        if(shdr->sh_type != SHT_PROGBITS && shdr->sh_type != SHT_NOBITS) continue;
+        if((shdr->sh_flags & SHF_ALLOC) == 0) continue;
+
+        uintptr_t base = shdr->sh_addr;
+        size_t limit = shdr->sh_size;
+
+        uintptr_t page_base = base & ~0xFFFULL;
+        size_t pml3i = (page_base >> PMLI_SHL(3)) & 0x1FF;
+        size_t pml2i = (page_base >> PMLI_SHL(2)) & 0x1FF;
+        size_t pml1i = (page_base >> PMLI_SHL(1)) & 0x1FF;
+        size_t base_pml0i = (page_base >> PMLI_SHL(0)) & 0x1FF;
+
+        pml4[pml3i].value = init_program_pml3_pptr | PAGE_PRESENT | PAGE_WRITEABLE | PAGE_USER;
+        s_init_program_pml3[pml2i].value = init_program_pml2_pptr | PAGE_PRESENT | PAGE_WRITEABLE | PAGE_USER;
+        s_init_program_pml2[pml1i].value = init_program_pml1_pptr | PAGE_PRESENT | PAGE_WRITEABLE | PAGE_USER;
+
+        size_t pages = (limit + 4095) / 4096;
+
+        for(size_t pmlo = 0; pmlo < pages; pmlo++) {
+            size_t pml0i = base_pml0i + pmlo;
+            if(pml0i > INIT_PROGRAM_PAGES_MAX) {
+                kpanic("init: Program exceeds maximum program pages in the same hugepage frame!\n");
+            }
+
+            uintptr_t pml0_pptr = init_program_pml0_pptr + (pml0i << 12);
+            s_init_program_pml1[pml0i].value = pml0_pptr | PAGE_PRESENT | PAGE_USER | PAGE_WRITEABLE;
+        }
+
+        //Copy data from section if SHT_PROGBITS
+        if(shdr->sh_type == SHT_PROGBITS) {
+            memcpy((void*)shdr->sh_addr, init_data + shdr->sh_offset, shdr->sh_size);
+        }
+    }
+
+    //Map the stack
+    pml4[255].value = init_stack_pml3_pptr | PAGE_PRESENT | PAGE_WRITEABLE | PAGE_USER | PAGE_XD;
+    s_init_stack_pml3[511].value = init_stack_pml2_pptr | PAGE_PRESENT | PAGE_WRITEABLE | PAGE_USER | PAGE_XD;
+    s_init_stack_pml2[511].value = init_stack_pml1_pptr | PAGE_PRESENT | PAGE_WRITEABLE | PAGE_USER | PAGE_XD;
+    for(size_t pmlo = 0; pmlo < INIT_STACK_PAGES; pmlo++) {
+        size_t pml0i = 511 - pmlo;
+        uintptr_t page_pptr = init_stack_pml0_pptr + (pmlo << 12);
+
+        s_init_stack_pml1[pml0i].value = page_pptr | PAGE_PRESENT | PAGE_WRITEABLE | PAGE_USER | PAGE_XD;
+    }
+
+    obj_release(init_pagedir);
+
+    //Enable syscalls
+    extern void _syscall_entry(void);
+    msr_set_lstar(_syscall_entry);
+
+    uintptr_t usp = 0x00007FFFFFFFFFF8;
+    uintptr_t uip = init_ehdr->e_entry;
+
+    __asm__ volatile("movq %0, %%rsp; \
+                      movq %1, %%rcx; \
+                      movq $0x202, %%r11; \
+                      swapgs; \
+                      sysretq"::
+                     "r"(usp), "r"(uip): "memory");
 }