fix: 6.12 kernel support and disable SIMD

- Disable SIMD by adding -mgeneral-regs-only compiler flag
- Fix paging_init hook: use bl instruction hook instead of direct function hook
- Add vmalloc_noprof symbol support for 6.12 kernel
- Add memset wrapper function
- Fix copy_to_user function fallback order
- Improve map_data stack allocation
- Update kptools to locate paging_init bl instruction in setup_arch

Author: NepXIN

kernel/Makefile

@@ -9,7 +9,7 @@ LD = $(TARGET_COMPILE)ld
 AS = $(TARGET_COMPILE)as
 OBJCOPY = $(TARGET_COMPILE)objcopy
 
-CFLAGS += -Wall -fno-builtin -std=gnu11 -nostdinc
+CFLAGS += -Wall -fno-builtin -std=gnu11 -nostdinc -mgeneral-regs-only
 CFLAGS += -g
 
 ifdef DEBUG

kernel/base/baselib.c

@@ -345,3 +345,8 @@ char *lib_strstr(const char *haystack, const char *needle)
 {
     return (char *)lib_memmem(haystack, lib_strlen(haystack), needle, lib_strlen(needle));
 }
+
+void *memset(void *dst, int c, size_t n)
+{
+    return lib_memset(dst, c, n);
+}

kernel/base/map.c

@@ -66,9 +66,9 @@ static void flush_icache_all(void)
     asm volatile("isb" : : : "memory");
 }
 
-static map_data_t *mem_proc()
+static void mem_proc(map_data_t *data)
 {
-    map_data_t *data = get_data();
+	*data = *get_data();
     uint64_t kernel_va = get_kva();
 
     // relocation
@@ -105,8 +105,6 @@ static map_data_t *mem_proc()
     uint64_t detect_virt = (uint64_t)((memblock_virt_alloc_try_nid_f)data->map_symbol.memblock_virt_alloc_relo)(
         0, 0x10, detect_phys, detect_phys, NUMA_NO_NODE);
     data->linear_voffset = detect_virt - detect_phys;
-
-    return data;
 }
 
 // todo: 52-bits pa
@@ -175,7 +173,10 @@ static uint64_t __noinline get_or_create_pte(map_data_t *data, uint64_t va, uint
 // todo: bti
 void __noinline _paging_init()
 {
-    map_data_t *data = mem_proc();
+	map_data_t buf;
+	map_data_t *data = &buf;
+    mem_proc(data);
+
 #ifdef MAP_DEBUG
     printk_f printk = (printk_f)(data->printk_relo);
 #define map_debug(idx, val) printk(data->str_fmt_px, idx, val)
@@ -202,10 +203,7 @@ void __noinline _paging_init()
         ((memblock_mark_nomap_f)(data->map_symbol.memblock_mark_nomap_relo))(start_pa, all_size);
 
     // paging_init
-    uint64_t paging_init_va = data->paging_init_relo;
-    *(uint32_t *)(paging_init_va) = data->paging_init_backup;
-    flush_icache_all();
-    ((paging_init_f)(paging_init_va))();
+    ((paging_init_f)(data->paging_init_relo))();
     // can't write data below
 
     // AttrIndx[2:0] encoding
@@ -243,7 +241,7 @@ void __noinline _paging_init()
     }
 
     flush_icache_all();
-
+	
     // free old start
     ((memblock_free_f)data->map_symbol.memblock_free_relo)(old_start_pa, reserve_size);
 

kernel/base/setup.h

@@ -25,8 +25,6 @@
 typedef struct
 {
     // preset
-    uint32_t paging_init_backup;
-    uint32_t __;
     int64_t map_offset;
     int64_t start_offset;
     int64_t start_size;
@@ -49,8 +47,7 @@ typedef struct
     uint64_t linear_voffset;
 } map_data_t;
 #else
-#define map_paging_init_backup_offset 0
-#define map_map_offset_offset (map_paging_init_backup_offset + 8)
+#define map_map_offset_offset 0
 #define map_start_offset_offset (map_map_offset_offset + 8)
 #define map_start_size_offset (map_start_offset_offset + 8)
 #define map_start_img_size_offset (map_start_size_offset + 8)

kernel/base/setup1.S

@@ -182,7 +182,10 @@ map_prepare:
     // map_data.paging_init_relo = setup_preset.paging_init_offset;
     ldr x11, [x10, #setup_paging_init_offset_offset]
     str x11, [x9, #map_paging_init_relo_offset]
-    mov x15, x11
+
+    ldr x15, [x10, #setup_paging_init_bl_offset_offset]
+    // uint64_t paging_init_bl_pa = paging_init_bl_offset + kernel_pa;
+    add x13, x15, x19
 
     // map_data.map_symbol = setup_preset.map_symbol
     add x0, x9, #map_map_symbol_offset
@@ -218,37 +221,6 @@ map_prepare:
     add x11, x11, #MEMORY_RW_SIZE
     str x11, [x9, #map_alloc_size_offset]
 
-    // backup and hook paging_init
-    // uint64_t paging_init_pa = paging_init_offset + kernel_pa;
-    add x13, x15, x19
-    // map_data.paging_init_backup = *(uint32_t *)(paging_init_pa);
-    ldr w12, [x13]
-
-    mov w3, #0x201F
-    movk w3, #0xD503, lsl#16
-    orr w1, w3, #0x100
-    mov w2, #0xFFFFFD1F
-    and w0, w12, w2
-    // if ((map_data.paging_init_backup & 0xFFFFFD1F) == 0xD503211F)
-    cmp w0, w1
-    b.ne .backup
-    // map_data.paging_init_backup = NOP
-    mov w12, w3
-    // uint32_t *p = (uint32_t *)paging_init_pa + 1;
-    add x11, x13, #4
-.cmp_auti:
-    // while ((*p & 0xFFFFFD1F) != 0xD503211F) ++p;
-    ldr w0, [x11], #4
-    and w0, w0, w2
-    cmp w0, w1
-    b.ne .cmp_auti
-    // *p = NOP
-    stur w3, [x11, #-4]
-
-.backup:
-    str w12, [x9, #map_paging_init_backup_offset]
-    dsb ish
-
     // uint64_t replace_offset = (uint64_t)(_paging_init - _map_start) + map_offset;
     adrp x11, _paging_init
     add x11, x11, :lo12:_paging_init
@@ -257,11 +229,11 @@ map_prepare:
     sub x11, x11, x12
     add x11, x11, x14
 
-    // *(uint32_t *)paging_init_pa = B_REL(paging_init_offset, replace_offset);
-    // #define B_REL(src, dst) (0x14000000u | (((dst - src) & 0x0FFFFFFFu) >> 2u))
+    // *(uint32_t *)paging_init_bl_pa = B_REL(paging_init_bl_offset, replace_offset);
+    // #define BL_REL(src, dst) (0x94000000u | (((dst - src) & 0x0FFFFFFFu) >> 2u))
     sub x15, x11, x15
     ubfx w15, w15, #2, #26
-    mov w12, #0x14000000
+    mov w12, #0x94000000
     orr w15, w15, w12
     str w15, [x13]
 

kernel/include/preset.h

@@ -236,6 +236,7 @@ typedef struct _setup_preset_t
     int64_t map_max_size;
     int64_t kallsyms_lookup_name_offset;
     int64_t paging_init_offset;
+    int64_t paging_init_bl_offset;
     int64_t printk_offset;
     map_symbol_t map_symbol;
     uint8_t header_backup[HDR_BACKUP_SIZE];
@@ -258,7 +259,8 @@ typedef struct _setup_preset_t
 #define setup_map_max_size_offset (setup_map_offset_offset + 8)
 #define setup_kallsyms_lookup_name_offset_offset (setup_map_max_size_offset + 8)
 #define setup_paging_init_offset_offset (setup_kallsyms_lookup_name_offset_offset + 8)
-#define setup_printk_offset_offset (setup_paging_init_offset_offset + 8)
+#define setup_paging_init_bl_offset_offset (setup_paging_init_offset_offset + 8)
+#define setup_printk_offset_offset (setup_paging_init_bl_offset_offset + 8)
 #define setup_map_symbol_offset (setup_printk_offset_offset + 8)
 #define setup_header_backup_offset (setup_map_symbol_offset + MAP_SYMBOL_SIZE)
 #define setup_superkey_offset (setup_header_backup_offset + HDR_BACKUP_SIZE)

kernel/linux/include/linux/vmalloc.h

@@ -64,6 +64,7 @@ extern void *kfunc_def(vm_map_ram)(struct page **pages, unsigned int count, int
 extern void kfunc_def(vm_unmap_aliases)(void);
 
 extern void *kfunc_def(vmalloc)(unsigned long size);
+extern void *kfunc_def(vmalloc_noprof)(unsigned long size);
 extern void *kfunc_def(vzalloc)(unsigned long size);
 extern void *kfunc_def(vmalloc_user)(unsigned long size);
 extern void *kfunc_def(vmalloc_node)(unsigned long size, int node);
@@ -121,6 +122,7 @@ static inline void vm_unmap_aliases(void)
 static inline void *vmalloc(unsigned long size)
 {
     kfunc_call(vmalloc, size);
+    kfunc_call(vmalloc_noprof, size);
     kfunc_not_found();
     return 0;
 }

kernel/patch/common/utils.c

@@ -75,15 +75,15 @@ int __must_check compat_copy_to_user(void __user *to, const void *from, int n)
 {
     int cplen = 0;
 
-    if (kfunc(seq_buf_to_user)) {
-        cplen = seq_buf_copy_to_user(to, from, n);
-    } else if (kfunc(xt_data_to_user)) {
+    if (kfunc(xt_data_to_user)) {
         // xt_data_to_user, xt_obj_to_user
         cplen = compat_xt_data_copy_to_user(to, from, n);
         if (!cplen) cplen = n;
     } else if (kfunc(bits_to_user)) {
         // bits_to_user, str_to_user
         cplen = compat_bits_copy_to_user(to, from, n);
+    } else if (kfunc(seq_buf_to_user)) {
+        cplen = seq_buf_copy_to_user(to, from, n);
     } else if (kfunc(trace_seq_to_user)) {
         cplen = trace_seq_copy_to_user(to, from, n);
     } else {

kernel/patch/ksyms/misc.c

@@ -336,6 +336,7 @@ void *kfunc_def(vm_map_ram)(struct page **pages, unsigned int count, int node) =
 void kfunc_def(vm_unmap_aliases)(void) = 0;
 
 void *kfunc_def(vmalloc)(unsigned long size) = 0;
+void *kfunc_def(vmalloc_noprof)(unsigned long size) = 0;
 void *kfunc_def(vzalloc)(unsigned long size) = 0;
 void *kfunc_def(vmalloc_user)(unsigned long size) = 0;
 void *kfunc_def(vmalloc_node)(unsigned long size, int node) = 0;
@@ -383,6 +384,7 @@ static void _linux_mm_vmalloc_sym_match(const char *name, unsigned long addr)
     // kfunc_match(vm_unmap_aliases, name, addr);
 
     kfunc_match(vmalloc, name, addr);
+    kfunc_match(vmalloc_noprof, name, addr);
     kfunc_match(vzalloc, name, addr);
     // kfunc_match(vmalloc_user, name, addr);
     // kfunc_match(vmalloc_node, name, addr);

tools/common.h

@@ -44,6 +44,7 @@ extern bool log_enable;
 #define align_ceil(x, align) (((uint64_t)(x) + (uint64_t)(align)-1) & ~((uint64_t)(align)-1))
 
 #define INSN_IS_B(inst) (((inst) & 0xFC000000) == 0x14000000)
+#define INSN_IS_BL(inst) (((inst) & 0xFC000000) == 0x94000000)
 
 #define bits32(n, high, low) ((uint32_t)((n) << (31u - (high))) >> (31u - (high) + (low)))