[parisc-linux] Mysterious hangs with parisc (a send_group_sig_info() analysis)

[Michael S. Zick]

On Sat July 22 2006 12:23, Carlos O'Donell wrote:
> > And here are a lot of questions:
> >         Q2: some diff in routine like 2 times ldcw with gcc-3 and only one with gcc-4 (may be ok?)
> >
> 
> Analyze the code. Is it doing functionally the same thing?
> 
> I assert that yes, it's doing exactly the same thing. The older gcc
> has unwrapped one of the ldcw's, and thus has two copies of the inner
> loop in the assembly. The code on the right does exactly the same
> thing, but in less instructions. GCC has gotten better.
> 
> There isn't a missing lock, we only needed one lock, the older gcc was
> doing a poorer job of code generation.
> 

OK - will go with that theory.

> There may be other parts of the kernel that have problems, so keep looking!
> 

Now that portion of the work is easy (not quick, easy) to reproduce.

Pick a fairly recent kernel (I think Joel has used both 2.6.16 & 2.6.17);
Use gcc-3.3 to build 64-bit-smp;
Use gcc-4.1 to build 64-bit-smp;

Pick a machine with two or more processors (Joel has PA8000 & PA8700);
Pick yourself a never-ending script that keeps all processors busy;

Run the gcc-4.1 built kernel; observe that machine locks up after some
time period ranging from hours to a few days;

Boot the gcc-3.3 built kernel; observe that machine runs for weeks.

Yes, there is something getting compiled differently than intended.

Q) Is the following an expected difference?

send_group_sig_info:                              send_group_sig_info:
    .PROC                                             .PROC
    .CALLINFO FRAME=144,CALLS,SAVE_RP,ENTRY_GR=6 |    .CALLINFO FRAME=160,CALLS,SAVE_RP,ENTRY_GR=7

Mike