Re: [DynInst_API:] Telling DynInst a particular function is non-returning

[Thomas Dullien <thomasdullien@xxxxxxxxxx>]

Hey there,

one quick question: After backporting the 3 commits, I tested with both recursive-mode disassembly and non-recursive mode

disassembly, and neither of the two worked. My question is now: Is this even supposed to work in non-recursive mode? E.g.

if we are not recursing into a called function, we can't tell whether it returns or not, right?

Cheers,

Thomas

On Thu, Mar 15, 2018 at 2:23 PM, Thomas Dullien <thomasdullien@xxxxxxxxxx> wrote:

Hey there,

ah, cool :-). I backported the 3 commits into the 9.3.2 codebase, but it does not seem to solve the issue just yet.

I will debug a bit more and report back once I understand what is going on.

Cheers,

Thomas

On Thu, Mar 15, 2018 at 2:09 PM, Xiaozhu Meng <xmeng@xxxxxxxxxxx> wrote:

Ah, I have a few pending commits related to this issue. Please look at my pull request at https://github.com/dyninst/dyninst/pull/437.

Not all commits are relevant to you, butÂbbac557,Â78b1bd1,Âda50a37Âshould be relevant.

For parsing_printf, it is "DYNINST_DEBUG_PARSING".

On Thu, Mar 15, 2018 at 8:02 AM, Thomas Dullien <thomasdullien@xxxxxxxxxx> wrote:

Hey there,

ok, rebuilding DynInst 9.3.2 now to investigate why adding the string does not seem to have any effect.

I looked at the source code, and I *think* the function is already in the list of non-returning functions.

Checking what is going on. Example setup:

.plt.got:0000000000053CF8 __stack_chk_fail proc near       ; CODE XREF: init:loc_54673âp

.plt.got:0000000000053CF8Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â; uninit:loc_54705âp ...

.plt.got:0000000000053CF8        Âjmp  Âcs:__stack_chk_fail_ptr

.plt.got:0000000000053CF8 __stack_chk_fail endp

.text:00000000000A4290 var_8Â Â Â Â Â Â= qword ptr -8

.text:00000000000A4290

.text:00000000000A4290 graph = rdi              Â; AVFilterGraph_0 *

.text:00000000000A4290 flags = rsi              Â; unsigned int

.text:00000000000A4290 ; __unwind {

.text:00000000000A4290        Âpush  rax

.text:00000000000A4291        Âmov  Ârax, fs:28h

.text:00000000000A429A        Âmov  Â[rsp+8+var_8], rax

.text:00000000000A429E        Âmov  Â[graph+5Ch], esi

.text:00000000000A42A1        Âmov  Ârax, fs:28h

.text:00000000000A42AA        Âcmp  Ârax, [rsp+8+var_8]

.text:00000000000A42AE        Âjnz  Âshort loc_A42B2

.text:00000000000A42B0        Âpop  Ârax

.text:00000000000A42B1Â Â Â Â Â Â Â Â Âretn

.text:00000000000A42B2 ; ---------------------------------------------------------------------------

.text:00000000000A42B2

.text:00000000000A42B2 loc_A42B2:Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â ; CODE XREF: avfilter_graph_set_auto_convert+1Eâj

.text:00000000000A42B2        Âcall  __stack_chk_fail

.text:00000000000A42B2 ; } // starts at A4290

.text:00000000000A42B2 avfilter_graph_set_auto_convert endp

DynInst for some reason does not interpret the tailing call as non-return. Digging to see what is going on,

will update here as I learn more :)

Cheers,

Thomas

On Thu, Mar 15, 2018 at 1:58 PM, Xiaozhu Meng <mxz297@xxxxxxxxx> wrote:

Hi Thomas,

On Thu, Mar 15, 2018 at 4:06 AM, Thomas Dullien <thomasdullien@xxxxxxxxxx> wrote:

Hey there,

ok, I have looked at a few options on how to best tackle this, and would love to solicit advice :-)

- I tried SymtabCodeSource::addNonReturning("__stack_chk_fail"); this did not seem to have an effect.

If you callÂSymtabCodeSource::addNonReturning("__stack_chk_fail") before calling CodeObject::parse(), this should work.Â

Â

- Looked at set_retstatus -- but that implies that the code is already parsed?

You are right that after CodeObject::parse() has finished, calling set_retstatus will only change the flag of return status for this function, will not re-parse the binary. So, we can focus on whyÂSymtabCodeSource::addNonReturning("__stack_chk_fail") does not work.Â

Â

Cheers,

Thomas

On Thu, Mar 15, 2018 at 9:28 AM, Thomas Dullien <thomasdullien@xxxxxxxxxx> wrote:

Ah. No. I just fund set_retstatus :-) -- please ignore my question :-)

On Thu, Mar 15, 2018 at 9:26 AM, Thomas Dullien <thomasdullien@xxxxxxxxxx> wrote:

Hey there,

after having my coffee, I realized: The proper way to do this is to derive from CodeSource

and overload the nonReturning functions, I guess? :)

Cheers,

Thomas

On Thu, Mar 15, 2018 at 9:14 AM, Thomas Dullien <thomasdullien@xxxxxxxxxx> wrote:

Hey there,

I am running into troubles with disassembling executables generated by

clang.3.8.1-24, for x64, with optimization set to size-optimize and stack cookies

enabled.

The trouble is that any function with an enabled stack cookie will end with a sequence

of:

 Epilogue to check stack cookie

 jnz .fail

 Rest of epilogue.

 retn

.fail:

 call __stack_checkfail  Â// Does not return

This leads to DynInst lumping all consecutive functions that use stack cookies

into one huge CFG.

Is there a way to tell DynInst that a particular function is not returning? I found

that the parseAPI allows me to query if a function returns, but I did not find anything

to "override" this behavior?

Cheers,

Thomas

_______________________________________________
Dyninst-api mailing list
Dyninst-api@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/dyninst-api

_______________________________________________
Dyninst-api mailing list
Dyninst-api@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/dyninst-api