arm-linux stub interfacing for lzma

src/p_vmlinx.cpp

@@ -770,6 +770,7 @@ void PackVmlinuxARM::pack(OutputFile *fo)
     cconf.conf_lzma.max_num_probs = 1846 + (768 << 4); // ushort: ~28KB stack
     compressWithFilters(&ft, 512, &cconf, getStrategy(ft));
     unsigned const txt_c_len = ph.c_len;
+    unsigned const txt_u_len = ph.u_len;
 
     const unsigned lsize = getLoaderSize();
 
@@ -777,8 +778,6 @@ void PackVmlinuxARM::pack(OutputFile *fo)
     if (0x40==(0xf0 & ft.id)) {
         linker->defineSymbol("filter_length", ph.u_len); // redefine
     }
-    linker->defineSymbol("BYTE_ADJ", (3& -txt_c_len));
-    linker->defineSymbol("WORD_ADJ", 4);
     defineDecompressorSymbols();
     relocateLoader();
 
@@ -794,24 +793,27 @@ void PackVmlinuxARM::pack(OutputFile *fo)
     shdro[1].sh_type = Elf32_Shdr::SHT_PROGBITS;
     shdro[1].sh_flags = Elf32_Shdr::SHF_ALLOC | Elf32_Shdr::SHF_EXECINSTR;
     shdro[1].sh_offset = fo_off;
-    shdro[1].sh_size = sizeof(stub_arm_linux_kernel_vmlinux_head) + ph.c_len + lsize;
+    shdro[1].sh_size = sizeof(stub_arm_linux_kernel_vmlinux_head) + txt_c_len + lsize;
     shdro[1].sh_addralign = 1;
 
     // This ought to be a linker section (to handle the relocation of 'bl'),
     // but buildLoader gets called from the middle of compressWithFilters
     // so there is a circularity problem.
-    fo->write(&stub_arm_linux_kernel_vmlinux_head[0], sizeof(stub_arm_linux_kernel_vmlinux_head)-2*4);
+    fo->write(&stub_arm_linux_kernel_vmlinux_head[0], sizeof(stub_arm_linux_kernel_vmlinux_head)-9*4);
-    tmp_le32 = get_le32(&stub_arm_linux_kernel_vmlinux_head[sizeof(stub_arm_linux_kernel_vmlinux_head)-2*4]);
+    tmp_le32 = get_le32(&stub_arm_linux_kernel_vmlinux_head[sizeof(stub_arm_linux_kernel_vmlinux_head)-9*4]);
-    tmp_le32 = (0xff000000 & tmp_le32) | (0x00ffffff & (-0+ ((3+ txt_c_len)>>2)));
+    tmp_le32 = (0xff000000 & tmp_le32) | (0x00ffffff & (7+ ((3+ txt_c_len)>>2)));
     fo->write(&tmp_le32, 4);
-    fo->write(&stub_arm_linux_kernel_vmlinux_head[sizeof(stub_arm_linux_kernel_vmlinux_head)-4], 4);
+    fo->write(&stub_arm_linux_kernel_vmlinux_head[sizeof(stub_arm_linux_kernel_vmlinux_head)-8*4], 5*4);
+    tmp_le32 = txt_c_len; fo->write(&tmp_le32, 4);
+    tmp_le32 = txt_u_len; fo->write(&tmp_le32, 4);
+    tmp_le32 = ph.method; fo->write(&tmp_le32, 4);
     fo_off += sizeof(stub_arm_linux_kernel_vmlinux_head);
 
     fo->write(obuf, ~3& (3+ txt_c_len)); fo_off += ~3& (3+ txt_c_len);
     fo->write(loader, lsize); fo_off += lsize;
 
 #if 0
-    printf("%-13s: compressed   : %8u bytes\n", getName(), ph.c_len);
+    printf("%-13s: compressed   : %8u bytes\n", getName(), txt_c_len);
     printf("%-13s: decompressor : %8u bytes\n", getName(), lsize);
 #endif
     verifyOverlappingDecompression();

src/stub/arm-linux.kernel.vmlinux-head.h

@@ -1,5 +1,5 @@
 /* arm-linux.kernel.vmlinux-head.h
-   created from arm-linux.kernel.vmlinux-head.bin, 20 (0x14) bytes
+   created from arm-linux.kernel.vmlinux-head.bin, 68 (0x44) bytes
 
    This file is part of the UPX executable compressor.
 
@@ -28,11 +28,14 @@
  */
 
 
-#define STUB_ARM_LINUX_KERNEL_VMLINUX_HEAD_SIZE    20
+#define STUB_ARM_LINUX_KERNEL_VMLINUX_HEAD_SIZE    68
-#define STUB_ARM_LINUX_KERNEL_VMLINUX_HEAD_ADLER32 0x621b0b3f
+#define STUB_ARM_LINUX_KERNEL_VMLINUX_HEAD_ADLER32 0xd30518ac
-#define STUB_ARM_LINUX_KERNEL_VMLINUX_HEAD_CRC32   0x87e74167
+#define STUB_ARM_LINUX_KERNEL_VMLINUX_HEAD_CRC32   0xa10e8134
 
-unsigned char stub_arm_linux_kernel_vmlinux_head[20] = {
+unsigned char stub_arm_linux_kernel_vmlinux_head[68] = {
-112,  0, 32,225,  1, 96, 34,233,  2,208,160,225,254,255,255,235,   /* 0x   0 */
+  9, 96, 34,233,  2,208,160,225,  0, 32,160,225, 36,  0,143,226,   /* 0x   0 */
-  2,160,157,232                                                    /* 0x  10 */
+ 10, 16,176,232,  4, 48, 45,229, 13, 48,160,225,  4,192, 45,229,   /* 0x  10 */
+254,255,255,235,  0,  0, 80,227,  7,  0,  0, 26, 14,112,157,232,   /* 0x  20 */
+  3,  0,130,224, 14,240,160,225,  0,  0,  0,  0,  0,  0,  0,  0,   /* 0x  30 */
+  0,  0,  0,  0                                                    /* 0x  40 */
 };

src/stub/src/arm-linux.kernel.vmlinux-head.S

@@ -28,12 +28,27 @@
 /* Calling sequence of equivalent code in arch/arm/boot/compressed/misc.c:
 decompress_kernel:  # (char *out, char *tmp, char *end_tmp, int arch_id)
 */
-        bkpt
+        stmdb r2!,{r0,r3,sp,lr}  // &outdata, arch_id, sp_in, retaddr
-        stmdb r2!,{r0,sp,lr}
+        mov sp,r2  // switch stacks to end_tmp area (64KB)
-        mov sp,r2
+        mov r2,r0  // &outdata
-// PackVmlinuxARM::pack knows the format of the next instruction.
+        adr r0,.L10
-        bl COMPRESSED_LENGTH
+        ldmia r0!,{r1,r3,ip}  // r0=&indata; r1=insize; r3=outsize; ip=method
-        ldmia sp,{r1,sp,pc}
+        str r3,[sp,#-4]!  // outsize
+        mov r3,sp  // &outsize
+        str ip,[sp,#-4]!  // method
+// PackVmlinuxARM::pack knows the layout of the following words.
+        bl LINUX000  // (&indata, insize, &outdata, &outsize, method)
+spin:
+        cmp r0,#0  // check for success
+        bne spin
+
+        ldmia sp,{r1,r2,r3,ip,sp,lr}  // method, outsize, &outdata, arch_id, sp_in, retaddr
+        add r0,r2,r3  // rv= &outdata[outsize]
+        mov pc,lr  // return
+.L10:
+        .long   COMPRESSED_LENGTH
+        .long UNCOMPRESSED_LENGTH
+        .long METHOD
 
 // Compressed data appears >here<, then decompressor.
 

src/stub/src/arm-linux.kernel.vmlinux.S

@@ -37,29 +37,18 @@
 ; ============= ENTRY POINT
 ; =============
 
-;  In:
-;    r0= outptr; r1= &tmp; r3= arch_id; lr= retaddr, inptr - N
-
   How to debug: run under qemu (http://fabrice.bellard.free.fr/qemu/)
   after un-commenting the  bkpt  opcode below.  That opcode forces qemu
   to stop in gdb.  You'll have to "set $pc+=4" by hand.
 */
 section         LINUX000
-        bkpt  // qemu breakpoint
+        // bkpt  // qemu DEBUG only
-        ldr r1,[lr,#-4]  // 'bl' instruction
-        str r3,[sp,$-4]!  // push arch_id
-        bic r1,r1,#~0<<24  // word count
-        sub sp,sp,#4  // space for outsize
-        mov r1,r1,lsl #2  // byte count
-        mov r3,sp  // &outsize
-        .long 0xe2411000 + BYTE_ADJ  // sub r1,r1,#BYTE_ADJ  // insize
-        mov r2,r0  // outptr
-        .long 0xe28e0000 + WORD_ADJ  // add r0,lr,#WORD_ADJ  // inptr
 /*
   r0= inptr
   r1= insize
   r2= outptr
   r3= &outsize
+  sp/ method
 */
 
 section         LXCALLT1