Merge pull request #14 from devkitPro/dynamic-parsing

Linkscript overhaul & .dynamic section parsing

buildscripts/lib/switch.ld

@@ -6,137 +6,157 @@ PHDRS
 	code   PT_LOAD FLAGS(5) /* Read | Execute */;
 	rodata PT_LOAD FLAGS(4) /* Read */;
 	data   PT_LOAD FLAGS(6) /* Read | Write */;
+	dyn    PT_DYNAMIC;
 }
 
 SECTIONS
 {
 	/* =========== CODE section =========== */
-
 	PROVIDE(__start__ = 0x0);
 	. = __start__;
 
-	.text ALIGN(0x1000) :
+	.crt0 :
+	{
+		KEEP (*(.crt0))
+		. = ALIGN(8);
+	} :code
+
+	.init :
 	{
-		/* .init */
-		KEEP( *(.crt0) )
 		KEEP( *(.init) )
-		. = ALIGN(4);
+		. = ALIGN(8);
+	} :code
 
-		/* .text */
+	.plt :
-		*(.text)
+	{
-		*(.text.*)
+		*(.plt)
-		*(.glue_7)
+		*(.iplt)
-		*(.glue_7t)
+		. = ALIGN(8);
-		*(.stub)
+	} :code
-		*(.gnu.warning)
-		*(.gnu.linkonce.t*)
-		. = ALIGN(4);
 
-		/* .fini */
+	.text :
+	{
+		*(.text.unlikely .text.*_unlikely .text.unlikely.*)
+		*(.text.exit .text.exit.*)
+		*(.text.startup .text.startup.*)
+		*(.text.hot .text.hot.*)
+		*(.text .stub .text.* .gnu.linkonce.t.*)
+		. = ALIGN(8);
+	} :code
+
+	.fini :
+	{
 		KEEP( *(.fini) )
-		. = ALIGN(4);
+		. = ALIGN(8);
-	} : code
+	} :code
 
 	/* =========== RODATA section =========== */
+	. = ALIGN(0x1000);
 
-	.rodata ALIGN(0x1000) :
+	.rodata :
 	{
-		*(.rodata)
+		*(.rodata .rodata.* .gnu.linkonce.r.*)
-		*(.roda)
+		. = ALIGN(8);
-		*(.rodata.*)
+	} :rodata
-		*all.rodata*(*)
-		*(.gnu.linkonce.r*)
-		SORT(CONSTRUCTORS)
-		. = ALIGN(4);
-	} : rodata
 
-	.ARM.extab : { *(.ARM.extab* .gnu.linkonce.armextab.*) } : rodata
+	.eh_frame_hdr     : { *(.eh_frame_hdr) *(.eh_frame_entry .eh_frame_entry.*) } :rodata
-	__exidx_start = .;
+	.eh_frame         : ONLY_IF_RO { KEEP (*(.eh_frame)) *(.eh_frame.*) } :rodata
-	ARM.exidx : { *(.ARM.exidx* .gnu.linkonce.armexidx.*) } : rodata
+	.gcc_except_table : ONLY_IF_RO { *(.gcc_except_table .gcc_except_table.*) } :rodata
-	__exidx_end = .;
+	.gnu_extab        : ONLY_IF_RO { *(.gnu_extab*) } : rodata
+
+	.dynamic           : { *(.dynamic) } :rodata :dyn
+	.interp            : { *(.interp) } :rodata
+	.note.gnu.build-id : { *(.note.gnu.build-id) } :rodata
+	.hash              : { *(.hash) } :rodata
+	.gnu.hash          : { *(.gnu.hash) } :rodata
+	.gnu.version       : { *(.gnu.version) } :rodata
+	.gnu.version_d     : { *(.gnu.version_d) } :rodata
+	.gnu.version_r     : { *(.gnu.version_r) } :rodata
+	.dynsym            : { *(.dynsym) } :rodata
+	.dynstr            : { *(.dynstr) } :rodata
+	.rela.dyn          : { *(.rela.*) } :rodata
 
 	/* =========== DATA section =========== */
+	. = ALIGN(0x1000);
 
-	.data ALIGN(0x1000) :
+	.eh_frame : ONLY_IF_RW { KEEP (*(.eh_frame)) *(.eh_frame.*) } :data
-	{
+	.gcc_except_table : ONLY_IF_RW { *(.gcc_except_table .gcc_except_table.*) } :data
-		*(.data)
+	.gnu_extab : ONLY_IF_RW { *(.gnu_extab*) } : data
-		*(.data.*)
+	.exception_ranges : ONLY_IF_RW { *(.exception_ranges .exception_ranges*) } :data
-		*(.gnu.linkonce.d*)
-		CONSTRUCTORS
-		. = ALIGN(4);
-                __got_start__ = .;
-                *(.got)
-	        __got_end__ = .;
-	} : data
 
-
+	.tdata ALIGN(8) :
-	.tdata ALIGN(4) :
 	{
 		__tdata_lma = .;
-		*(.tdata)
+		*(.tdata .tdata.* .gnu.linkonce.td.*)
-		*(.tdata.*)
+		. = ALIGN(8);
-		*(.gnu.linkonce.td.*)
-		. = ALIGN(4);
 		__tdata_lma_end = .;
-	} : data
+	} :data
 
-	.tbss ALIGN(4) :
+	.tbss ALIGN(8) :
 	{
-		*(.tbss)
+		*(.tbss .tbss.* .gnu.linkonce.tb.*) *(.tcommon)
-		*(.tbss.*)
+		. = ALIGN(8);
-		*(.gnu.linkonce.tb.*)
+	} :data
-		*(.tcommon)
-		. = ALIGN(4);
-	} : data
 
-	.preinit_array ALIGN(4) :
+	.preinit_array ALIGN(8) :
 	{
 		PROVIDE (__preinit_array_start = .);
 		KEEP (*(.preinit_array))
 		PROVIDE (__preinit_array_end = .);
-	} : data
+	} :data
 
-	.init_array ALIGN(4) :
+	.init_array ALIGN(8) :
 	{
 		PROVIDE (__init_array_start = .);
 		KEEP (*(SORT(.init_array.*)))
 		KEEP (*(.init_array))
 		PROVIDE (__init_array_end = .);
-	} : data
+	} :data
 
-	.fini_array ALIGN(4) :
+	.fini_array ALIGN(8) :
 	{
 		PROVIDE (__fini_array_start = .);
 		KEEP (*(.fini_array))
 		KEEP (*(SORT(.fini_array.*)))
 		PROVIDE (__fini_array_end = .);
-	} : data
+	} :data
 
-	.ctors ALIGN(4) :
+	.ctors ALIGN(8) :
 	{
 		KEEP (*crtbegin.o(.ctors)) /* MUST be first -- GCC requires it */
 		KEEP (*(EXCLUDE_FILE (*crtend.o) .ctors))
 		KEEP (*(SORT(.ctors.*)))
 		KEEP (*(.ctors))
-	} : data
+	} :data
 
-	.dtors ALIGN(4) :
+	.dtors ALIGN(8) :
 	{
 		KEEP (*crtbegin.o(.dtors))
 		KEEP (*(EXCLUDE_FILE (*crtend.o) .dtors))
 		KEEP (*(SORT(.dtors.*)))
 		KEEP (*(.dtors))
-	} : data
+	} :data
+
+	__got_start__ = .;
+
+	.got            : { *(.got) *(.igot) } :data
+	.got.plt        : { *(.got.plt)  *(.igot.plt) } :data
+
+	__got_end__ = .;
+
+	.data ALIGN(8) :
+	{
+		*(.data .data.* .gnu.linkonce.d.*)
+		SORT(CONSTRUCTORS)
+	} :data
 
 	__bss_start__ = .;
-	.bss ALIGN(4) :
+	.bss ALIGN(8) :
 	{
 		*(.dynbss)
-		*(.bss)
+		*(.bss .bss.* .gnu.linkonce.b.*)
-		*(.bss.*)
-		*(.gnu.linkonce.b*)
 		*(COMMON)
-		. = ALIGN(4);
+		. = ALIGN(8);
 
 		/* Reserve space for the TLS segment of the main thread */
 		__tls_start = .;

buildscripts/lib/switch.specs

@@ -1,7 +1,7 @@
 %rename link                old_link
 
 *link:
-%(old_link) -T switch.ld%s -d --gc-sections
+%(old_link) -T switch.ld%s -pie --gc-sections -z text
 
 *startfile:
 switch_crt0%O%s crti%O%s crtbegin%O%s

crt0/switch_crt0.s

@@ -3,55 +3,39 @@
 
 _start:
     bl startup
-
+    .ascii "HOMEBREW"
-    .word 0x454d4f48
-    .word 0x57455242
 
 startup:
     // get aslr base
     sub  x28, x30, #4
 
     // clear .bss
-    ldr  x0, =__bss_start__
+    adrp x0, __bss_start__
-    ldr  x1, =__bss_end__
+    adrp x1, __bss_end__
+    add  x0, x0, #:lo12:__bss_start__
+    add  x1, x1, #:lo12:__bss_end__
     sub  x1, x1, x0  // calculate size
     add  x1, x1, #7  // round up to 8
     bic  x1, x1, #7
-    mov  x2, #0
-    add  x0, x0, x28 // relocate ptr
 
 bss_loop: 
-    str  x2, [x0], #8
+    str  xzr, [x0], #8
     subs x1, x1, #8
     bne  bss_loop
 
-    // relocate .got
+    // process .dynamic section
-    ldr  x0, =__got_start__
-    ldr  x1, =__got_end__
-    sub  x1, x1, x0  // calculate size
-    add  x1, x1, #7  // round up to 8
-    bic  x1, x1, #7
-    add  x0, x0, x28 // relocate ptr
-
-got_loop: 
-    ldr  x2, [x0]
-    add  x2, x2, x28
-    str  x2, [x0], #8
-    subs x1, x1, #8
-    bne  got_loop
-
     mov  x0, x28
-    ldr  x3, =initSystem
+    adrp x1, _DYNAMIC
-    add  x3, x3, x28
+    add  x1, x1, #:lo12:_DYNAMIC
-    blr  x3
+    bl   __nx_dynamic
 
+    // initialize system
+    mov  x0, x28
+    bl   __nx_init
+
+    // call entrypoint
     mov  x0, #0 // argc
     mov  x1, #0 // argv
-
+    adrp x30, __nx_exit
-    ldr  x3, =main
+    add  x30, x30, #:lo12:__nx_exit
-    add  x3, x3, x28
+    b    main
-    ldr  x2, =__nx_exit
-    add  x2, x2, x28
-    mov  x30, x2
-    br  x3
-

nx/Makefile

@@ -32,9 +32,9 @@ INCLUDES	:=	include
 #---------------------------------------------------------------------------------
 # options for code generation
 #---------------------------------------------------------------------------------
-ARCH	:=	-march=armv8-a
+ARCH	:=	-march=armv8-a -fPIC
 
-CFLAGS	:=	-g -Wall -Werror -fPIC \
+CFLAGS	:=	-g -Wall -Werror \
 			-ffunction-sections \
 			-fdata-sections \
 			$(ARCH) \

nx/source/system/dynamic.c

@@ -0,0 +1,37 @@
+#include <switch/types.h>
+#include <elf.h>
+
+void __nx_dynamic(uintptr_t base, const Elf64_Dyn* dyn)
+{
+	const Elf64_Rela* rela = NULL;
+	u64 relasz = 0;
+
+	for (; dyn->d_tag != DT_NULL; dyn++)
+	{
+		switch (dyn->d_tag)
+		{
+			case DT_RELA:
+				rela = (const Elf64_Rela*)(base + dyn->d_un.d_ptr);
+				break;
+			case DT_RELASZ:
+				relasz = dyn->d_un.d_val / sizeof(Elf64_Rela);
+				break;
+		}
+	}
+
+	if (rela == NULL)
+		for (;;); // Panic
+
+	for (; relasz--; rela++)
+	{
+		switch (ELF64_R_TYPE(rela->r_info))
+		{
+			case R_AARCH64_RELATIVE:
+			{
+				u64* ptr = (u64*)(base + rela->r_offset);
+				*ptr = base + rela->r_addend;
+				break;
+			}
+		}
+	}
+}

nx/source/system/sysinit.s

@@ -1,99 +1,19 @@
         .text
-	.global	initSystem
+	.global	__nx_init
-	.type	initSystem,	%function
+	.type	__nx_init, %function
 
-initSystem:
+__nx_init:
 	stp x29, x30, [sp, #-16]!
-        adr     x1, __nx_binarybase
-        str     x0, [x1]
 	bl	__libnx_init
-
 	bl	__appInit
-	bl	__nx_libc_init_array
+	bl	__libc_init_array
 	ldp x29, x30, [sp], #16
 	ret
 
 	.global	__nx_exit
-	.type	__nx_exit,	%function
+	.type	__nx_exit, %function
 
 __nx_exit:
-	bl	__nx_libc_fini_array
+	bl	__libc_fini_array
 	bl	__appExit
-
 	b	__libnx_exit
-
-__nx_libc_init_array:
-	stp x29, x30, [sp, #-16]!
-	stp x21, x22, [sp, #-16]!
-	stp x19, x20, [sp, #-16]!
-        adr     x3, __nx_binarybase
-        ldr     x20, [x3]
-
-        ldr     x0, =__preinit_array_start
-        ldr     x1, =__preinit_array_end
-        sub     x1, x1, x0
-        add     x21, x0, x20
-        lsr     x19, x1, #3
-        cbz     x19, __nx_libc_init_array_end0
-
-        __nx_libc_init_array_lp0:
-        ldr     x3, [x21], #8
-        sub     x19, x19, #1
-        add     x3, x3, x20
-        blr     x3
-        cbnz    x19, __nx_libc_init_array_lp0
-
-__nx_libc_init_array_end0:
-        bl      _init
-
-        ldr     x0, =__init_array_start
-        ldr     x1, =__init_array_end
-        sub     x1, x1, x0
-        add     x21, x0, x20
-        lsr     x19, x1, #3
-        cbz     x19, __nx_libc_init_array_end1
-
-        __nx_libc_init_array_lp1:
-        ldr     x3, [x21], #8
-        sub     x19, x19, #1
-        add     x3, x3, x20
-        blr     x3
-        cbnz    x19, __nx_libc_init_array_lp1
-
-__nx_libc_init_array_end1:
-	ldp x19, x20, [sp], #16
-	ldp x21, x22, [sp], #16
-	ldp x29, x30, [sp], #16
-        ret
-
-__nx_libc_fini_array:
-	stp x29, x30, [sp, #-16]!
-	stp x21, x22, [sp, #-16]!
-	stp x19, x20, [sp, #-16]!
-        adr     x3, __nx_binarybase
-        ldr     x20, [x3]
-
-        ldr     x0, =__fini_array_start
-        ldr     x1, =__fini_array_end
-        sub     x1, x1, x0
-        add     x21, x0, x20
-        lsr     x19, x1, #3
-        cbz     x19, __nx_libc_fini_array_end
-
-        __nx_libc_fini_array_lp:
-        sub     x19, x19, #1
-        ldr     x3, [x21, x19, lsl #3]
-        add     x3, x3, x20
-        blr     x3
-        cbnz    x19, __nx_libc_fini_array_lp
-
-__nx_libc_fini_array_end:
-	ldp x19, x20, [sp], #16
-	ldp x21, x22, [sp], #16
-	ldp x29, x30, [sp], #16
-        ret
-
-        .data
-__nx_binarybase:
-        .dword 0
-

tools/elf2nso.c

@@ -62,8 +62,8 @@ uint8_t* ReadEntireFile(const char* fn, size_t* len_out) {
 
 int main(int argc, char* argv[]) {
     if (argc != 3) {
-        printf("%s <elf-file> <nso-file>\n", argv[0]);
+        fprintf(stderr, "%s <elf-file> <nso-file>\n", argv[0]);
-        return 1;
+        return EXIT_FAILURE;
     }
 
     NsoHeader nso_hdr;
@@ -72,48 +72,57 @@ int main(int argc, char* argv[]) {
     nso_hdr.Unk3 = 0x3f;
 
     if (sizeof(NsoHeader) != 0x100) {
-        printf("Bad compile environment!\n");
+        fprintf(stderr, "Bad compile environment!\n");
-        return 1;
+        return EXIT_FAILURE;
     }
 
     size_t elf_len;
     uint8_t* elf = ReadEntireFile(argv[1], &elf_len);
     if (elf == NULL) {
-        printf("Failed to open input!\n");
+        fprintf(stderr, "Failed to open input!\n");
-        return 1;
+        return EXIT_FAILURE;
     }
 
     if (elf_len < sizeof(Elf64_Ehdr)) {
-        printf("Input file doesn't fit ELF header!\n");
+        fprintf(stderr, "Input file doesn't fit ELF header!\n");
-        return 1;
+        return EXIT_FAILURE;
     }
 
     Elf64_Ehdr* hdr = (Elf64_Ehdr*) elf;
     if (hdr->e_machine != EM_AARCH64) {
-        printf("Invalid ELF: expected AArch64!\n");
+        fprintf(stderr, "Invalid ELF: expected AArch64!\n");
-        return 1;
+        return EXIT_FAILURE;
     }
 
-    Elf64_Off ph_end = hdr->e_phoff + 3 * sizeof(Elf64_Phdr);
+    Elf64_Off ph_end = hdr->e_phoff + hdr->e_phnum * sizeof(Elf64_Phdr);
 
     if (ph_end < hdr->e_phoff || ph_end > elf_len) {
-        printf("Invalid ELF: phdrs outside file!\n");
+        fprintf(stderr, "Invalid ELF: phdrs outside file!\n");
-        return 1;
+        return EXIT_FAILURE;
     }
 
-    if (hdr->e_phnum != 3) {
+    Elf64_Phdr* phdrs = (Elf64_Phdr*) &elf[hdr->e_phoff];
-        printf("Invalid ELF: expected 3 phdrs!\n");
+    size_t i, j = 0;
-        return 1;
-    }
-
-    Elf64_Phdr* phdr = (Elf64_Phdr*) &elf[hdr->e_phoff];
-    size_t i;
     size_t file_off = sizeof(NsoHeader);
 
     uint8_t* comp_buf[3];
     int comp_sz[3];
 
-    for (i=0; i<3; i++, phdr++) {
+    for (i=0; i<3; i++) {
+        Elf64_Phdr* phdr = NULL;
+        while (j < hdr->e_phnum) {
+            Elf64_Phdr* cur = &phdrs[j++];
+            if (cur->p_type == PT_LOAD) {
+                phdr = cur;
+                break;
+            }
+        }
+
+        if (phdr == NULL) {
+            fprintf(stderr, "Invalid ELF: expected 3 loadable phdrs!\n");
+            return EXIT_FAILURE;
+        }
+
         nso_hdr.Segments[i].FileOff = file_off;
         nso_hdr.Segments[i].DstOff = phdr->p_vaddr;
         nso_hdr.Segments[i].DecompSz = phdr->p_filesz;
@@ -133,16 +142,16 @@ int main(int argc, char* argv[]) {
         comp_buf[i] = malloc(comp_max);
 
         if (comp_buf[i] == NULL) {
-            printf("Compressing: Out of memory!\n");
+            fprintf(stderr, "Compressing: Out of memory!\n");
-            return 1;
+            return EXIT_FAILURE;
         }
 
         // TODO check p_offset
         comp_sz[i] = LZ4_compress_default(&elf[phdr->p_offset], comp_buf[i], phdr->p_filesz, comp_max);
 
         if (comp_sz[i] < 0) {
-            printf("Failed to compress!\n");
+            fprintf(stderr, "Failed to compress!\n");
-            return 1;
+            return EXIT_FAILURE;
         }
 
         nso_hdr.CompSz[i] = comp_sz[i];
@@ -152,8 +161,8 @@ int main(int argc, char* argv[]) {
     FILE* out = fopen(argv[2], "wb");
 
     if (out == NULL) {
-        printf("Failed to open output file!\n");
+        fprintf(stderr, "Failed to open output file!\n");
-        return 1;
+        return EXIT_FAILURE;
     }
 
     // TODO check retvals
@@ -162,5 +171,5 @@ int main(int argc, char* argv[]) {
     for (i=0; i<3; i++)
         fwrite(comp_buf[i], comp_sz[i], 1, out);
 
-    return 0;
+    return EXIT_SUCCESS;
 }