I'm not sure why we haven't seen this before, but I started seeing a
stack trace on reboot on my rx2600 ia64 box running latest 2.6. Here it
is:
[<a0000001006ba020>] get_boot_pages+0x0/0x2c0
sp=e00000010267fc80 bsp=e000000102679130
[<a0000001004a8fb0>] sba_alloc_coherent+0x70/0x1a0
sp=e00000010267fc80 bsp=e0000001026790f8
[<a00000010047e3b0>] mptscsih_synchronize_cache+0x1d0/0x640
sp=e00000010267fc80 bsp=e000000102679010
[<a0000001004665e0>] mptbase_shutdown+0xc0/0xe0
sp=e00000010267fd30 bsp=e000000102678fd8
[<a0000001003a9140>] device_shutdown+0x260/0x280
sp=e00000010267fd30 bsp=e000000102678fa8
[<a0000001000b10a0>] sys_reboot+0x2e0/0x720
sp=e00000010267fd30 bsp=e000000102678f50
sba_alloc_coherent is simply calling __get_free_pages() to setup a DMA
mapping for the mpt driver to sync a disk. However, we've already
cleared system_running in sys_reboot, so w/ CONFIG_NUMA, we blowup
trying to call get_boot_pages, which was already freed. I'm not sure if
the proper fix is to make get_boot_pages not an __init function or if
the CONFIG_NUMA code really intends to be keying off something else.
Thoughts on the right fix? Thanks,
Alex
--
Alex Williamson HP Linux & Open Source Lab