WIP: [Bolt] Add support for DT_INIT_ARRAY

[kaadam]

Currently Bolt relies on ELF 'e_entry' field or DT_INIT to determine entry point of an ELF file for the instrumentation. There is a case when an ELF file only has DT_INIT_ARRAY/DT_FINI_ARRAY sections, and ELF 'e_entry' holds zero.

[kaadam]

TODO(s):

• Add test case for DT_INIT_ARRAY
• Add runtime test for DT_INIT_ARRAY

[zhouyu0502]

hello i have come across a similar question .
I want to bolt instrument for libc.so.6 (aarch64)
it seems success but when run bin with intrumented libc.so.6, it did not exe _bolt_instr_start() which is in e_entry
i think it is because my libc.so.6 have no DT_INIT and have HasInterpHeader at the same time
how can i fix this ?

[zhouyu0502]

hello i have applyed for libc.so.6 (have no DT_INIT),but i still failed collect prof because libc.so.6 HasInterpHeader ,SO IN THIS CASE how can i fixed?

[zhouyu0502]

i have applyed your pr but still failed when using intrumented libc.so.6 in arm64.it seems the father process exited after __bolt_instr_setup and can not exe __bolt_start_trampoline ,i do not know why

[kaadam]

Hi, thanks for your comment and your findings.

Glibc is a tricky Elf shared object. It can acts as a binary or as a shared object.
So libc.so has an entry point which points to '__libc_main' make it directly runnable. It just prints a banner.
File from csu/version.c

extern void __libc_main (void) __attribute__ ((noreturn));

void __libc_main (void)
{
  __libc_print_version ();
  _exit (0);
}

The draft solution what I uploaded, distinguishes between whether the ELF file is a binary or a shared object based on presence of the Interp header.
In this case libc is an so, and it has Interp header, so Bolt just deals with Elf e_entry point which is the __libc_main, and hooking its start runtime function to that. However this never runs if libc acts as SO.

So most probably DT_INITARRAY should be used here (if no DT_INIT), instead of e_entry.
I need to look at this example.

[zhouyu0502]

Hi, thanks for your comment and your findings.

Glibc is a tricky Elf shared object. It can acts as a binary or as a shared object. So libc.so has an entry point which points to '__libc_main' make it directly runnable. It just prints a banner. File from csu/version.c

extern void __libc_main (void) __attribute__ ((noreturn));

void __libc_main (void)
{
  __libc_print_version ();
  _exit (0);
}

The draft solution what I uploaded, distinguishes between whether the ELF file is a binary or a shared object based on presence of the Interp header. In this case libc is an so, and it has Interp header, so Bolt just deals with Elf e_entry point which is the __libc_main, and hooking its start runtime function to that. However this never runs if libc acts as SO.

So most probably DT_INITARRAY should be used here (if no DT_INIT), instead of e_entry. I need to look at this example.

in my case when i apply your pr i have deleted !BC->HasInterpHeader ,
case ELF::DT_INIT_ARRAY
BC->InitArrayAddress = Dyn.getPtr();
break;
case ELF::DT_INIT_ARRAYSZ:
BC->InitArraySize = Dyn.getPtr();

and i found it still cannot exe bin (who use libc.so.6) successfully but i check that it exe __bolt_instr_setup and can not exe __bolt_start_trampoline but when exe _start in libc.so.6 it failed
i dont know why ,thanks for your reply
and i will show more details,wait for me

[zhouyu0502]

1. my proposeis to instrument libc.so.6 (arm64) so not bin
2. i found thatmy origin libc.so.6 do not have INIT (which means cannot exe __boIt_instr_start(),but it have INIT_ARRAY、INIT ARRAYSZ, and PT_INTERP,in order to fix this i have applyed your pr and delete
   (!BC->HaslnterpHeader) in boIt/lib/Rewrite/Rewritelnstance.cpp in case
   DT_INIT and DT_INIT_ARRAY
3. howerver when i use instrumented libc.so.6 in a exe it still failed to work properliy, i think some went wrong after exe
   bolt_start_trampoline (which means jump to the _start)
   4.feel very confused and want to debug this,can vou tell me how to fix this bug?

[zhouyu0502]

hi ,sorry to interrupt you ,
i attach some information in case you need
1.file libc.so.6
libc.so.6: ELF 64-bit LSB shared object, ARM aarch64, version 1 (GNU/Linux), dynamically linked, interpreter /lib/ld-linux-aarch64.so.1, BuildID[sha1]=f5e52c2f342cbaafd253543336b5e0416b7f2b02, for GNU/Linux 3.14.0, with debug_info, not stripped
2.readelf -l libc.so.6
Elf file type is DYN (Shared object file)
Entry point 0x2b330
There are 10 program headers, starting at offset 64
......
INTERP 0x000000000015d198 0x000000000015d198 0x000000000015d198
0x000000000000001b 0x000000000000001b R 0x8
[Requesting program interpreter: /lib/ld-linux-aarch64.so.1]
......
3.readelf -d libc.so.6 | grep INIT
0x0000000000000019 (INIT_ARRAY) 0x19c970
0x000000000000001b (INIT_ARRAYSZ) 24 (bytes)

How