[Cbe-oss-dev] CELL SDK 2.1's dmabecnh + 2.6.23-rc3 (geoff/ps3-linux.git) = ps3 hangs

[Goffredo Marocchi]

I am running Fedora 7 and interestingly enough, before compiling the
kernel from Geoff Levand's git repository, with the default
2.6.22.x.fc7 kernel  the dmabench benchmark included in the 2.1
release of the CELL SDK would SEGFAULT.

With the freshly compiled 2.6.23-rc3 kernel I get the following result:

# cat /etc/redhat-release
Fedora release 7 (Moonshine)

# uname -ra
Linux IGGS-PS3 2.6.23-rc3-g966fd9c0 #1 SMP Sat Aug 25 10:49:23 CEST
2007 ppc64 ppc64 ppc64 GNU/Linux

[root at IGGS-PS3 dma]# ./dmabench --numspes 1 seqdmar
Time base frequency measured as = 79.801 MHz
All SPEs completed successfully!
dmabench results: seqdmar numspes=1 numreqs=1 entrysize=128
   dma_size         ticks       pclocks     microsecs     aggr GB/s
------------  ------------  ------------   -----------  ------------
          8          11.5           461          0.14        0.0555
         16          11.5           461          0.14        0.1110
         32          11.5           461          0.14        0.2221
         64          11.3           453          0.14        0.4512
        128          11.6           463          0.14        0.8844
        256          12.0           480          0.15        1.7053
        512          12.7           510          0.16        3.2096
       1024          14.5           581          0.18        5.6356
       2048          17.9           715          0.22        9.1558
       4096          27.5          1102          0.34       11.8902
       8192          47.3          1895          0.59       13.8267
      16384          86.7          3475          1.09       15.0871

It no longer SEGFAULTs, but after running this program again (doing
some read and write tests too), the PS3 hangs (crash, stops
responding, etc...) after running this test:

./dmabench --numspes 5 seqdmar

Seeing how it works well with the stock 2.6.16 kernel in FC-6 (as
confirmed by IBM's devs on the developerworks forum) which is the one
they tested the 2.1 CELL SDK with, since it SEGFAULTs with the stock
2.6.22 F-7 kernel and it works again in most cases with the custom
(from Geoff's tree) 2.6.23-rc3 kernel although bringing the whole
system down (which brings back memories of wrongly directed DMA
transfers on PS2Linux ;)) in some instances... well, I wonder if there
is something that needs to be addressed on the kernel side.

If you need any more data or logs, please do ask and I will try to
provide them asap.

Best Regards,

Goffredo Marocchi