Re: [PATCH] ARC: prevent showing irrelevant exception info in signal message

[Vineet Gupta]

+CC Al

On 06/29/2018 12:39 PM, Eugeniy Paltsev wrote:
> We process signals in the end of syscall/exception handler.
> It the signal is fatal we print register's content using
> show_regs function. show_regs() also prints information about
> last exception happened.
>
> In case of multicore system we can catch the situation when we
> will print wrong information about exception. See the example:
> ______________________________
> CPU-0: started to handle page fault
> CPU-1: sent signal to process, which is executed on CPU-0
> CPU-0: ended page fault handle. Started to process signal before
>        returnig to userspace. Process signal, which is send from
>        CPU-0. As th signal is fatal we call show_regs().
>        show_regs() will show information about last exception
>        which is *page fault* (instead of "trap" which is used for
>        signals and happened on CPU-0)
>
> So we will get message like this:
>     /home/waitpid02
>   potentially unexpected fatal signal 8.
>   Path: /home/waitpid02
>   CPU: 0 PID: 100 Comm: waitpid02 Not tainted 4.10.0-rc4 #2
>   task: 9f11c200 task.stack: 9f3ae000
>
>   [ECR   ]: 0x00050200 => Invalid Write @ 0x00000000 by insn @ 0x000123ec
>   [EFA   ]: 0x00000000
>   [BLINK ]: 0x123ea
>   [ERET  ]: 0x123ec
>     @off 0x123ec in [/home/waitpid02]
>     VMA: 0x00010000 to 0x00016000
>   [STAT32]: 0x80080882 : IE U
>   BTA: 0x000123ea  SP: 0x5ffd3db0  FP: 0x00000000
>   LPS: 0x20031684 LPE: 0x2003169a LPC: 0x00000006
>   [-----other-info-----]
>
> This message is confusing because it show information about page fault
> ( [ECR   ]: 0x00050200 => Invalid Write ) which is absolutely irrelevant
> to signal.

Agreed this is misleading. @Al, is there a way to identify process termination
from signals because it did something wrong vs. say unhandled signal. For former,
we want to dump additional info in show_regs() such as PC / Fault addres etc and
not in other scenario.

> This situation was reproduced with waitpid02 LTP test.
> _____________________________
>
> So remove printing information about exceptions from show_regs()
> to avoid confusing messages. Print information about exceptions
> only in required places instead of show_regs()
>
> Now we don't print information about exceptions if signal is simply
> send by another userspace app. So in case of waitpid02 we will print
> next message:
> _____________________________
>     ./waitpid02
>   potentially unexpected fatal signal 8.
>   [STAT32]: 0x80080082 : IE U
>   BTA: 0x20000fc4	 SP: 0x5ff8bd64	 FP: 0x00000000
>   LPS: 0x200524a0	LPE: 0x200524b6	LPC: 0x00000006
>   [-----other-info-----]
> _____________________________

The prints I'm seeing now, for a segv from NULL pointer access is even more
confusing !
There's a mixup of prints....

-------------------->8--------------------
Path: /segv
CPU: 0 PID: 70 Comm: segv Not tainted 4.17.0+ #412

[ECRÂÂ ]: 0x00050200 => Invalid Write @ 0x00000000 by insn @ 0x000103ac
[EFAÂÂ ]: 0x00000000
[BLINK ]: 0x20047bb0
[ERETÂ ]: 0x103ac
ÂÂÂ @off 0x103ac in [/segv]
ÂÂÂ VMA: 0x00010000 to 0x00012000

potentially unexpected fatal signal 11.
[STAT32]: 0x80080882 : IE UÂÂÂÂ
BTA: 0x00010398ÂÂÂ ÂSP: 0x5fc95e1cÂÂÂ ÂFP: 0x5fc95e20
LPS: 0x20039ffcÂÂÂ LPE: 0x2003a000ÂÂÂ LPC: 0x00000000
r00: 0x00000001ÂÂÂ r01: 0x5fc95e94ÂÂÂ r02: 0x00000000ÂÂÂ
r03: 0x00000064ÂÂÂ r04: 0x80808080ÂÂÂ r05: 0x2f2f2f2fÂÂÂ
...
-------------------->8--------------------

and for the process killed by signal 8, we get below.

-------------------->8--------------------
[ARCLinux]# kill -8 71
[ARCLinux]# potentially unexpected fatal signal 8.
[STAT32]: 0x80080882 : IE UÂÂÂÂ
BTA: 0x20020660ÂÂÂ ÂSP: 0x5fbcddecÂÂÂ ÂFP: 0x5fbcde1c
LPS: 0x20039ffcÂÂÂ LPE: 0x2003a000ÂÂÂ LPC: 0x00000000
r00: 0xfffffdfcÂÂÂ r01: 0x5fbcddf0ÂÂÂ r02: 0x00000000ÂÂÂ
r03: 0x00000008ÂÂÂ r04: 0x80808080ÂÂÂ r05: 0x2f2f2f2fÂÂÂ
r06: 0x7a2f5f4aÂÂÂ r07: 0x00000000ÂÂÂ r08: 0x00000065ÂÂÂ
...


[1]+Â Floating point exceptionÂÂ ./sleep
-------------------->8--------------------

I'm not sure whats the improvement here vs. the status quo.

For signal based kill, we don't want to dump the extra registers and if any, we
might still want to print the PC where the process was last seen in user mode to
give user of idea what it was doing at the time.

-Vineet