Re: [Dri-devel] Mesa SIGSEGV's due to broken common_x86_asm.S

[Brian P. <br...@tu...>]

Brian Paul wrote:
> Linus Torvalds wrote:
> 
>> Ok,  I'm a bit bitter, because I just spent a long time chasing down a 
>> kernel bug that didn't turn out to be a kernel bug at all.
>>
>> I started seeing that strange SIGSEGV with programs that use dri, and it
>> happened right after the SIGFPE that tested for XMM support. As it
>> happens, I've done some signal delivery changes in the kernel lately, 
>> so I
>> blamed myself.
>>
>> It wasn't my fault.
>>
>> That frigging Mesa assembly-code is broken. In particular, it undoes the
>> whole stack frame _before_ it does the divide-by-zero thing, yet it still
>> has stuff in the local frame. Signal delivery will overwrite the local
>> frame if the stack is aligned just the right way, and as a result the
>> LDMXCSR that follows the DIVPS will load crap into MXCSR. And the crap it
>> loads may well cause a SIGSEGV due to a GP-fault by the CPU.
>>
>> Grr. Double-grr. That assembly-language is written in some unreadable 
>> syntax anyway, but here's a totally untested diff that may fix the 
>> crap by de-allocating the stack only after we're actually _done_ with it.
>>
>> I was too lazy to check whether the other asm routines were similarly 
>> broken. Anyway, the rule is:
>>  YOU MUST NOT USE LOCATIONS ON THE STACK UNDER THE STACK POINTER
>>
>> Ok, I feel better now after that rant. Sorry,
> 
> 
> Thanks for the patch.  I've checked in the change to all the relevant 
> trees.
>  From the CVS logs, it looks like this problem was introduced by a patch 
> from June 2002.

Josh Vanderhoof informed me that 16, not 32, should be added to ESP at the end 
of that routine.  I've checked in this change.

-Brian