gdbserver crash on mips - Illegal instruction

[patryk4815]

isa			: mips1 mips2 mips32r1 mips32r2
ASEs implemented	: mips16
Options implemented	: tlb 4kex 4k_cache 32fpr prefetch mcheck ejtag llsc dc_aliases perf_cntr_intr_bit mm_full

[patryk4815]

Coredump:
gdbserver.1743875995.2223.4.core.zip

Binary from here: https://github.com/guyush1/gdb-static/releases/download/v16.2-static/gdb-static-mips.tar.gz

[patryk4815]

Crash on FPU instruction.

LLM response:
The root cause:

• The program you're running (likely gdbserver) was compiled with FPU instructions (e.g., sdc1, add.d, mov.d, etc.), which are not supported by your platform.
• Your system is using soft-float (-msoft-float), meaning no hardware support for floating point operations.

[guyush1]

Hmmm...
I had to do some guesswork, since gdbserver is compiled with a symbol-stripping option (i.e -s).

However, i am fairly certain that gdbserver called longjmp.
The reason for the crash seems to be because your cpu does not support 64-bit floating point operations.
This code emitted can be seen in libc's source code:

It seems like a solution for your problem would be to use a different toolchain - one that comes with a libc that is compiled without this opcodes enabled. Unfortunately, simply passing -msoft-float isn't enough - since libc.a / libc.so comes pre-compiled...

[guyush1]

Btw: if you could conpile gdbserver without the -s flag (simply remove it from build.sh), and then reproduce the crash - it would be helpful to make sure we are indeed crashing on longjmp

[patryk4815]

I just manually compiled gdb+static musl with float=soft and it is working fine

[guyush1]

Musl is good since its source code is better and probably does not contain the same problem

[guyush1]

Btw, did you compile just gdbserver in static mode via musl? I tried changing our build scrupts to use musl instead of glibc to solve similar problems, but had problems with libexpat / libgmp that were tough to resolve, so it would be helpful to know if you managed to compile the whole thing via musl

[patryk4815]

I just used nix:
nix build nixpkgs#pkgsCross.mips-linux-gnu.pkgsStatic.gdb

But I had to fix bug in gdb manually: https://sourceware.org/bugzilla/show_bug.cgi?id=21070
sed -i '[email protected]@asm/sgidefs.h@g' ./gdb/mips-linux-nat.c

[patryk4815]

rust-lang/rust#34910 (comment)
🤔 looks like most of mips devices have only soft float?

[guyush1]

@patryk4815
This should be fixed in our latest release now:
https://github.com/guyush1/gdb-static/releases/tag/v16.3-static

If anything still does not work, please reopen this :)

[patryk4815]

@guyush1 still same issue,

root@OpenWrt:~# /tmp/gdbserver 0:1234 /bin/sh
Illegal instruction

Crash at 0x21b04 offset

[guyush1]

Hello, i will take a look at this soon-ish. It is probably related to setjmp yet again and can be resolved via using the proper compiler

[guyush1]

Hello @patryk4815 , I think i have a solution for you.

I compiled gdbserver with a soft-float mips compiler. This should guarantee that the problematic inline assembly at longjmp (which is not effected by -msoft-float) is not present anymore, so no floating point operations will be performed in longjmp and cause you to crash.

Please try it and tell me if it works :).

gdbserver-mips-soft.zip

[patryk4815]

Thanks works! :)

[guyush1]

Closing this since we found a solution. I might add a mips-sf (soft-float) variant if more people need it.