285 lines
14 KiB
C
285 lines
14 KiB
C
#ifndef UTSM_LINUX_LOADER_H
|
||
#define UTSM_LINUX_LOADER_H
|
||
|
||
#include <utsm/types.h>
|
||
|
||
/* ===== Linux boot protocol structures (Documentation/x86/boot.rst) ===== */
|
||
|
||
/* bzImage setup_header — lives at offset 0x1f1 inside bzImage and boot_params.
|
||
* Only the fields we actually use are listed; gaps are reserved/padding. */
|
||
struct linux_setup_header {
|
||
u8 setup_sects; /* 0x1f1: number of setup sectors */
|
||
u16 root_flags; /* 0x1f2 */
|
||
u32 syssize; /* 0x1f4: protected-mode code size / 16 */
|
||
u16 ram_size; /* 0x1f8 */
|
||
u16 vid_mode; /* 0x1fa */
|
||
u16 boot_flag; /* 0x1fe: 0xAA55 */
|
||
u16 jump; /* 0x200: jump insn */
|
||
u32 header; /* 0x202: "HdrS" magic */
|
||
u16 version; /* 0x206: boot protocol version */
|
||
u32 realmode_swtch; /* 0x208 */
|
||
u16 start_sys_seg; /* 0x20c */
|
||
u16 kernel_version; /* 0x20e: offset to version string */
|
||
u8 type_of_loader; /* 0x210 */
|
||
u8 loadflags; /* 0x211 */
|
||
u16 setup_move_size; /* 0x212 */
|
||
u32 code32_start; /* 0x214: 32-bit entry (also 64-bit entry base) */
|
||
u32 ramdisk_image; /* 0x218 */
|
||
u32 ramdisk_size; /* 0x21c */
|
||
u32 bootsect_kludge; /* 0x220 */
|
||
u16 heap_end_ptr; /* 0x224 */
|
||
u8 ext_loader_ver; /* 0x226 */
|
||
u8 ext_loader_type; /* 0x227 */
|
||
u32 cmd_line_ptr; /* 0x228: command line physical address */
|
||
u32 init_size; /* 0x22c: total init memory needed */
|
||
u32 acpi_rsdp_addr; /* 0x230 */
|
||
u8 _pad0[4]; /* 0x234: relocatable_kernel + min_alignment */
|
||
u16 xloadflags; /* 0x236: bit0=XLF_KERNEL_64 */
|
||
u32 cmdline_size; /* 0x238 */
|
||
u32 _pad1; /* 0x23c */
|
||
u64 setup_data; /* 0x240: linked list of setup_data */
|
||
u64 pref_address; /* 0x248 → wait, this is payload_offset */
|
||
|
||
/* NOTE: The real layout continues with payload_offset at 0x248.
|
||
* However, pref_address is at 0x258 and init_size at 0x260.
|
||
* The struct above is intentionally simplified — we read fields
|
||
* by offset rather than relying on this struct for the full layout. */
|
||
} __attribute__((packed));
|
||
|
||
/* boot_params is 4096 bytes. The setup_header is at offset 0x1f1. */
|
||
#define BOOT_PARAMS_SIZE 4096
|
||
#define SETUP_HEADER_OFFSET 0x1f1
|
||
#define BOOT_PARAMS_ACPI_RSDP_ADDR_OFF 0x070
|
||
#define BOOT_PARAMS_RSDP_COPY_OFF 0x800
|
||
#define BOOT_PARAMS_RSDP_COPY_SIZE 36
|
||
|
||
/* Key field offsets within bzImage / boot_params */
|
||
#define BZ_BOOT_FLAG_OFF 0x1fe
|
||
#define BZ_HEADER_MAGIC 0x53726448 /* "HdrS" */
|
||
#define BZ_HEADER_OFF 0x202
|
||
#define BZ_VERSION_OFF 0x206
|
||
#define BZ_TYPE_OF_LOADER_OFF 0x210
|
||
#define BZ_LOADFLAGS_OFF 0x211
|
||
#define BZ_CODE32_START_OFF 0x214
|
||
#define BZ_RAMDISK_IMAGE_OFF 0x218
|
||
#define BZ_RAMDISK_SIZE_OFF 0x21c
|
||
#define BZ_CMD_LINE_PTR_OFF 0x228
|
||
#define BZ_INIT_SIZE_OFF 0x22c /* wait — actually init_size is at 0x260 */
|
||
#define BZ_XLOADFLAGS_OFF 0x236
|
||
#define BZ_CMDLINE_SIZE_OFF 0x238
|
||
#define BZ_PAYLOAD_OFFSET_OFF 0x248
|
||
#define BZ_PAYLOAD_LENGTH_OFF 0x24c
|
||
#define BZ_PREF_ADDRESS_OFF 0x258
|
||
#define BZ_INIT_SIZE_ACTUAL_OFF 0x260
|
||
|
||
/* xloadflags bits */
|
||
#define XLF_KERNEL_64 0x0001
|
||
#define XLF_CAN_BE_LOADED_ABOVE 0x0002
|
||
#define XLF_EFI_HANDOVER 0x0010
|
||
#define XLF_EFI_KEXEC 0x0020
|
||
|
||
/* loadflags bits */
|
||
#define LOADFLAG_LOADED_HIGH 0x01
|
||
#define LOADFLAG_KEEP_ROOT 0x02
|
||
#define LOADFLAG_CAN_USE_HEAP 0x80
|
||
|
||
/* e820 entry types */
|
||
#define E820_TYPE_RAM 1
|
||
#define E820_TYPE_RESERVED 2
|
||
#define E820_TYPE_ACPI 3
|
||
#define E820_TYPE_NVS 4
|
||
#define E820_TYPE_UNUSABLE 5
|
||
#define E820_TYPE_PMEM 7
|
||
|
||
/* e820_table is at 0x2d0 in Linux 6.6 boot_params (arch/x86/include/uapi/asm/bootparam.h).
|
||
* e820_entries count is at 0x1e8. Each entry is 20 bytes: u64 addr, u64 size, u32 type. */
|
||
#define E820_ENTRIES_OFFSET 0x1e8
|
||
#define E820_TABLE_OFFSET 0x2d0
|
||
#define E820_MAX_ENTRIES 128
|
||
#define E820_ENTRY_SIZE 20
|
||
|
||
/* ===== Guest memory layout (GPA) =====
|
||
* Based on utsm-linux-dual-kernel-architecture.md.
|
||
* 0x00000000-0x003FFFFF : UTSM reserved (denied to Linux)
|
||
* 0x00400000-0x004FFFFF : boot_params / cmdline / GDT / page tables / stack
|
||
* 0x01000000-0x02FFFFFF : Linux kernel image
|
||
* 0x03000000-0x03FFFFFF : Linux initrd
|
||
* 0x04000000-0x040FFFFF : IPC shared memory (1MB, both kernels RW)
|
||
* 0x04100000-0x04FFFFFF : Graphics surface pool (16MB, VSCode integration Phase 1)
|
||
* 0x05000000-0x5FFFFFFF : Linux general RAM (~1.4GB, expanded Phase 1 for VSCode)
|
||
* 0xF4000000-0xF4007000 : virtio-mmio devices (8 slots, expanded Phase 1)
|
||
*/
|
||
#define LINUX_GUEST_KERNEL_GPA 0x01000000ULL /* 16 MB, 2MB-aligned */
|
||
#define LINUX_GUEST_BOOTPARAMS_GPA 0x00400000ULL /* 4 MB */
|
||
#define LINUX_GUEST_CMDLINE_GPA 0x00500000ULL /* 5 MB */
|
||
#define LINUX_GUEST_INITRD_GPA 0x03000000ULL /* 48 MB (within kernel+initrd region) */
|
||
#define LINUX_GUEST_STACK_GPA 0x00800000ULL /* 8 MB (stack grows down) */
|
||
#define LINUX_GUEST_GDT_GPA 0x00600000ULL /* 6 MB (guest GDT) */
|
||
#define LINUX_GUEST_PGT_GPA 0x00700000ULL /* 7 MB (guest page tables) */
|
||
#define LINUX_GUEST_CMDLINE_MAX 4096
|
||
|
||
/* IPC shared memory GPA (must match CODE/utsm-ipc/ipc_proto.h) */
|
||
#define LINUX_GUEST_IPC_SHM_GPA 0x04000000ULL /* 64 MB */
|
||
#define LINUX_GUEST_IPC_SHM_SIZE (1024 * 1024) /* 1MB */
|
||
|
||
/* VSCode integration Phase 1: Graphics surface pool.
|
||
* Separate DMA region EPT-mapped into the Linux guest. Linux user-space mmaps
|
||
* /dev/utsm at this GPA to obtain a writable scanout buffer that UTSM reads
|
||
* for blitting to the Deshab framebuffer (Phase 3 virtio-gpu shadow buffer).
|
||
* 16MB accommodates a 1920×1080×4 framebuffer (~8MB) with headroom for
|
||
* cursor plane and dirty rect tracking metadata. */
|
||
#define LINUX_GUEST_SURFACE_GPA 0x04100000ULL /* 65 MB (immediately after IPC shm) */
|
||
#define LINUX_GUEST_SURFACE_SIZE (16 * 1024 * 1024) /* 16MB */
|
||
|
||
/* General RAM starts after IPC + surface regions.
|
||
* Phase 1 expansion: 1.4GB to fit VSCode + Electron + GUI stack. */
|
||
#define LINUX_GUEST_RAM_GPA 0x05000000ULL /* 80 MB */
|
||
#define LINUX_GUEST_RAM_SIZE 0x5B000000ULL /* 1.4GB (0x05000000..0x5FFFFFFF) */
|
||
|
||
/* ===== virtio-mmio device GPA layout =====
|
||
* virtio-mmio 设备放在 GPA 高位 MMIO 区（guest RAM 之上，不占用 e820 RAM）。
|
||
* guest 通过 cmdline 参数 virtio_mmio.device= 发现这些设备。
|
||
* 每个设备占 0x200 字节寄存器空间（virtio-mmio 规范要求 4KB 对齐）。
|
||
*
|
||
* 这些 GPA 在 EPT 中【故意不映射】，访问时触发 EPT violation，
|
||
* 由 vmexit handler 路由到 virtio_mmio 后端模拟寄存器读写。
|
||
*
|
||
* GPA 选择 0xF4000000 起始（接近 4GB，远离 guest RAM 0x05000000-0x5FFFFFFF）。
|
||
*
|
||
* VSCode integration Phase 1 expansion: 槽位从 3 扩到 8，为后续阶段预留：
|
||
* slot 0: blk (Phase 1-) IRQ5
|
||
* slot 1: net (Phase 1-) IRQ6
|
||
* slot 2: rootfs (Phase 1-) IRQ7
|
||
* slot 3: keyboard (Phase 2) IRQ8 — virtio-input, evdev /dev/input/event0
|
||
* slot 4: mouse (Phase 2) IRQ9 — virtio-input, evdev /dev/input/event1
|
||
* slot 5: gpu (Phase 3) IRQ10 — virtio-gpu 2D, /dev/dri/card0
|
||
* slot 6: extra-rootfs (Phase 4) IRQ11 — persistent rw volume (/dev/vdc)
|
||
* slot 7: reserved — future use
|
||
*/
|
||
#define VIRTIO_MMIO_GPA_BASE 0xF4000000ULL
|
||
#define VIRTIO_MMIO_GPA_STRIDE 0x1000ULL /* 每个设备 4KB */
|
||
#define VIRTIO_MMIO_MAX_SLOTS 8 /* Phase 1: 3→8 */
|
||
|
||
#define VIRTIO_MMIO_BLK_GPA (VIRTIO_MMIO_GPA_BASE + 0 * VIRTIO_MMIO_GPA_STRIDE) /* 0xF4000000 */
|
||
#define VIRTIO_MMIO_NET_GPA (VIRTIO_MMIO_GPA_BASE + 1 * VIRTIO_MMIO_GPA_STRIDE) /* 0xF4001000 */
|
||
#define VIRTIO_MMIO_ROOTFS_GPA (VIRTIO_MMIO_GPA_BASE + 2 * VIRTIO_MMIO_GPA_STRIDE) /* 0xF4002000 */
|
||
/* VSCode integration: Phase 2/3/4 device slots (currently unused, reserved) */
|
||
#define VIRTIO_MMIO_KBD_GPA (VIRTIO_MMIO_GPA_BASE + 3 * VIRTIO_MMIO_GPA_STRIDE) /* 0xF4003000 */
|
||
#define VIRTIO_MMIO_MOUSE_GPA (VIRTIO_MMIO_GPA_BASE + 4 * VIRTIO_MMIO_GPA_STRIDE) /* 0xF4004000 */
|
||
#define VIRTIO_MMIO_GPU_GPA (VIRTIO_MMIO_GPA_BASE + 5 * VIRTIO_MMIO_GPA_STRIDE) /* 0xF4005000 */
|
||
#define VIRTIO_MMIO_EXTRA_ROOTFS_GPA (VIRTIO_MMIO_GPA_BASE + 6 * VIRTIO_MMIO_GPA_STRIDE) /* 0xF4006000 */
|
||
|
||
/* 判断 GPA 是否落在 virtio-mmio 设备区（含两端） */
|
||
#define VIRTIO_MMIO_GPA_END (VIRTIO_MMIO_GPA_BASE + VIRTIO_MMIO_MAX_SLOTS * VIRTIO_MMIO_GPA_STRIDE)
|
||
#define IS_VIRTIO_MMIO_GPA(gpa) ((gpa) >= VIRTIO_MMIO_GPA_BASE && (gpa) < VIRTIO_MMIO_GPA_END)
|
||
|
||
/* ===== Linux loader API ===== */
|
||
|
||
/* Loaded Linux guest state */
|
||
struct linux_guest_info {
|
||
u64 kernel_gpa; /* GPA where kernel was loaded */
|
||
u64 kernel_entry; /* 64-bit entry point GPA */
|
||
u64 kernel_size; /* Size of loaded kernel image */
|
||
u64 bootparams_gpa; /* GPA of boot_params */
|
||
u64 cmdline_gpa; /* GPA of command line */
|
||
u64 cmdline_size; /* Command line length (with NUL) */
|
||
u64 initrd_gpa; /* GPA of initrd (0 = none) */
|
||
u64 initrd_size; /* Size of initrd */
|
||
u64 pgt_gpa; /* GPA of guest page tables */
|
||
u64 gdt_gpa; /* GPA of guest GDT */
|
||
int loaded; /* 1 = successfully loaded */
|
||
/* VSCode integration Phase 1: graphics surface pool.
|
||
* Allocated as a separate DMA buffer and EPT-mapped into the guest at
|
||
* surface_gpa. Exposed to Linux user-space via UTSM_HCALL_SURFACE_INFO
|
||
* and /dev/utsm mmap (see CODE/linux/patches/utsm_hcall.c). */
|
||
u64 surface_hpa; /* Host physical address of surface pool (0 = none) */
|
||
u64 surface_gpa; /* GPA visible inside Linux guest */
|
||
u64 surface_size; /* Surface pool size in bytes */
|
||
};
|
||
|
||
/* Find the Linux bzImage boot module among Limine modules.
|
||
* Returns pointer to the module data, or NULL if not found.
|
||
* Sets *size_out to the module size. */
|
||
void *linux_find_bzimage_module(u64 *size_out);
|
||
|
||
/* Find the Arch rootfs image boot module (path contains "rootfs").
|
||
* Exposed to Linux guest via memory-backed virtio-blk as /dev/vdb. */
|
||
void *linux_find_rootfs_module(u64 *size_out);
|
||
|
||
/* VSCode Phase 4: find the extra rootfs image boot module (path contains
|
||
* "extra"). Exposed via writable virtio-blk slot6 as /dev/vdc (overlayfs
|
||
* upper / VSCode data volume). Returns NULL when the module is absent. */
|
||
void *linux_find_extra_rootfs_module(u64 *size_out);
|
||
|
||
/* VSCode Phase 5: find the VSCode tarball boot module (path contains
|
||
* "vscode", e.g. linux-vscode.tar.gz). The tarball is streamed into the
|
||
* guest and extracted to /opt/vscode by lxc_vscode_install().
|
||
* Returns NULL when the module is absent (install skipped, non-fatal). */
|
||
void *linux_find_vscode_module(u64 *size_out);
|
||
|
||
/* Parse bzImage header and return the protected-mode image's file offset,
|
||
* load length, decompression allocation size, and startup_64 entry offset.
|
||
* The protected-mode image starts after setup sectors; setup_header's
|
||
* payload_offset instead locates compressed data inside that image.
|
||
* Returns 0 on success, negative on error. */
|
||
int linux_parse_bzimage(const void *bzimage, u64 size,
|
||
u64 *kernel_offset_out, u64 *kernel_length_out,
|
||
u64 *init_size_out, u64 *entry_offset_out);
|
||
|
||
/* Initialize the Linux guest: parse bzImage, allocate guest memory,
|
||
* set up boot_params, page tables, GDT, and EPT mappings.
|
||
* Returns 0 on success, negative on error. */
|
||
int linux_loader_init(void);
|
||
|
||
/* Get the loaded Linux guest info (NULL if not loaded). */
|
||
const struct linux_guest_info *linux_get_guest_info(void);
|
||
|
||
/* VSCode integration Phase 1: query graphics surface pool.
|
||
* Returns 0 and fills outputs if the surface pool was allocated and EPT-mapped
|
||
* into the Linux guest; returns -1 if not initialized (e.g. Linux guest not
|
||
* loaded). hpa_out receives the host physical address (for UTSM-side blit),
|
||
* gpa_out receives the GPA visible inside Linux (for the guest mmap path),
|
||
* size_out receives the pool size in bytes. Any output pointer may be NULL. */
|
||
int linux_get_surface_info(u64 *hpa_out, u64 *gpa_out, u64 *size_out);
|
||
|
||
/* VSCode Phase 3: return host virtual address of the graphics surface pool.
|
||
* The virtio-gpu 2D backend composites the scanout into this buffer; the
|
||
* desktop reads it to blit guest graphics onto the IDE host area.
|
||
* Returns NULL if the pool was not allocated (Linux guest not loaded / alloc failed). */
|
||
void *linux_get_surface_vaddr(void);
|
||
|
||
/* Configure VMCS for Linux guest and perform vmlaunch.
|
||
* Must be called after vmm_init() + linux_loader_init().
|
||
* Returns 0 on successful guest termination, negative on error. */
|
||
int linux_launch(void);
|
||
|
||
/* Default Linux command line (can be overridden).
|
||
* Daemon PARK is a VMCALL. Guest idle may HLT; that stays in the
|
||
* guest (wait-for-IRQ). Do not use nohlt/idle=poll — poll idle
|
||
* NULL-derefs current after the first real sleep.
|
||
* noapic/nolapic/nosmp: UTSM VMM 只模拟 legacy PIC(8259) + PIT(8254)，
|
||
* 不提供 LAPIC/IOAPIC；guest 走 XT-PIC 模式，IRQ vector = 0x30+irq。
|
||
* virtio_mmio.device=: 向 guest 声明 virtio-mmio 设备（blk@0xF4000000 IRQ5,
|
||
* net@0xF4001000 IRQ6），需 CONFIG_VIRTIO_MMIO_CMDLINE_DEVICES=y。
|
||
* VSCode Phase 2: kbd@0xF4003000 IRQ8, mouse@0xF4004000 IRQ9 (virtio-input)。
|
||
* VSCode Phase 3: gpu@0xF4005000 IRQ10 (virtio-gpu 2D, /dev/dri/card0)。
|
||
* VSCode Phase 4: xrootfs@0xF4006000 IRQ11 (可写 virtio-blk, /dev/vdc)。
|
||
* 注：无 extra-rootfs boot module 时，后端改映射 ESP 上的
|
||
* boot/linux-extra-rootfs.img（AHCI+FAT32），避免 Limine 把 4GiB
|
||
* 打进 RAM。文件也没有则不注册，guest 探测超时跳过。 */
|
||
#define LINUX_DEFAULT_CMDLINE \
|
||
"earlyprintk=serial,ttyS0,115200 console=ttyS0,115200 " \
|
||
"nokaslr debug loglevel=8 noinvpcid nopcid nopti pti=off nogbpages " \
|
||
"tsc_early_khz=2000000 lpj=4000000 i8042.nokbd i8042.noaux " \
|
||
"no_timer_check " \
|
||
"noapic nolapic nosmp ibt=off user_shstk=off " \
|
||
"virtio_mmio.device=4K@0xF4000000:5 " \
|
||
"virtio_mmio.device=4K@0xF4001000:6 " \
|
||
"virtio_mmio.device=4K@0xF4002000:7 " \
|
||
"virtio_mmio.device=4K@0xF4003000:8 " \
|
||
"virtio_mmio.device=4K@0xF4004000:9 " \
|
||
"virtio_mmio.device=4K@0xF4005000:10 " \
|
||
"virtio_mmio.device=4K@0xF4006000:11"
|
||
|
||
#endif
|