Kernel panic when a program triggers a floating-point trap

[bhaible]

Find attached a program that

1) invokes feenableexcept(FE_INVALID); to turn FE_INVALID floating-point exceptions (not C++ exceptions!) into traps,

2) exercises an arithmetic operation with a signalling NaN, so as to trigger such a FE_INVALID floating-point exception.

How to reproduce:

• Open a Terminal.
• Unpack testdir4.tar.gz.
• Then

  cd testdir4
  ./configure
  make
  make check
  

Expected outcome: The program crashes with some signal (most likely SIGFPE).

Actual outcome: Kernel panic; screenshot attached.

The program invocation that crashes is one of:

gltests/test-snan-2 f
gltests/test-snan-2 d
gltests/test-snan-2 l

[bhaible]

test case

[bhaible]

screenshot of kernel panic

[bhaible]

More details:

gltests/test-snan-2 f succeeds.

gltests/test-snan-2 d succeeds.

gltests/test-snan-2 l is the invocation that crashes.

[pulkomandy]

the backtrace goes through x86_unexpected_exception, I think that means we are missing some code to handle these?

[waddlesplash]

What hrev, please?

Looks similar to #14802 which was solved by hrev55135.

[waddlesplash]

Actually it looks even more similar to #15781.

[bhaible]

What hrev, please?

I'm using the release from ​https://www.haiku-os.org/get-haiku/r1beta4/, download timestamp: 2022-12-22.

[X512]

I suspect that exception occurs when kernel write FPU registers during contex switch, some of value is SNaN and triggering exceptions by SNaN is enabled in FPU flags. Kernel should not trust FPU flags set by userland and turn off SNaN exceptions before context switch.

[korli]

​https://review.haiku-os.org/c/haiku/+/7089

[korli]

Fixed in hrev57363

[bhaible]

Thanks! I confirm it's fixed in hrev57370.