[Cbe-oss-dev] PATCH [0/7] spusched enhancements

[Luke Browning]

On Fri, 2007-11-02 at 17:49 +1100, Jeremy Kerr wrote:

> * These are based on the 2.6.22 tree + SDK3 patches. In particular, they 
> require Arnd's block-in-spu_acquire_runnable.patch for the ctx->run_wq 
> waitqueue. I'm in the process of re-rolling these for the spufs tree 
> (based on current powerpc head) though, and will post the updated 
> patches when I'm done. The most pervasive change is the transition from 
> SPUCTX_UTIL_* to SPU_UTIL_*.

I am not familiar with the SPU_UTIL_* changes.  When were they posted?

Will you integrate this patch set also?  I can get Andre to do it, if
you would prefer.  I am not going to have much time over the next month
or so, so I am going to try to limit my time to review and debug
issues. 


> * The patch "[4/7] decouple spu scheduler from spufs_spu_run" introduces 
> a bug for overcommitment. The consume-all test[1] blocks forever, and 
> results in an uninterruptable process. After kill -9-ing it, I end up 
> with contexts left hanging around, and a whole bunch of blocked kernel 
> threads:

Yes, these problems are fixed in patches 6 and 7.  Patches 0 and 1 are
standalone, I think 3 is also, but once you get to 4 you need all of
them. 

> 
> Call Trace:
> [c0000000020cf6d0] [c0000000020cf770] 0xc0000000020cf770 (unreliable)
> [c0000000020cf8a0] [c00000000000e9f0] .__switch_to+0xcc/0x100
> [c0000000020cf930] [c00000000036ef9c] .schedule+0x5c4/0x670
> [c0000000020cfa20] [c00000000036ff64] .__mutex_lock_slowpath+0x1c8/0x32c
> [c0000000020cfaf0] [d000000000107a0c] .spu_activate+0x304/0x6d4 [spufs]
> [c0000000020cfbd0] [d000000000109b28] .spufs_run_spu+0x668/0xd9c [spufs]
> [c0000000020cfcd0] [d00000000010568c] .do_spu_run+0xb4/0x17c [spufs]
> [c0000000020cfd80] [c000000000039b30] .sys_spu_run+0xb8/0x124
> [c0000000020cfe30] [c000000000007534] syscall_exit+0x0/0x40
> 
> I'll see if I can chase down the problem, but do you have any ideas 
> here?  

No.  Maybe, this means patch 3 is not standalone.