Re: iommu_alloc failure and panic

[Anton Blanchard <anton@samba.org>]

Hi,

> I have been testing large numbers of storage devices.  I have 16000 scsi
> LUNs configured.  (4000 fiberchannel LUNS seen 4 times).  They are
> configured as 4000 multipath devices using device mapper.  I have 100 of
> those devices configured as a logical volume using LVM2.  Once I start
> bonnie++ using one of those logical volumes I start seeing iommu_alloc
> failures and then a panic.  I don't have this issue when accessing the
> individual devices or the individual multipath devices.  Only when
> conglomerated into a logical volume.
> 
> Here is the syslog output:
> 
> Jan 26 14:56:41 linux kernel: iommu_alloc failed, tbl c00000000263c480 vaddr c0000000d7967000 npages 10

This stuff should be OK since the lpfc driver should handle the iommu
filling up. Im guessing since you have so many LUNs you can get a whole
lot of outstanding IO, enough to fill up the TCE table.

> DAR: C000000600711710

> NIP [C00000000000F7D0] .validate_sp+0x30/0xc0
> LR [C00000000000FA2C] .show_stack+0xec/0x1d0
> Call Trace:
> [C0000000076DBBC0] [C00000000000FA18] .show_stack+0xd8/0x1d0 (unreliable)
> [C0000000076DBC60] [C000000000433838] .schedule+0xd78/0xfb0
> [C0000000076DBDE0] [C000000000079FC0] .worker_thread+0x1b0/0x1c0

This is interesting, an oops in validate_sp which is a pretty boring
function. We have seen this in the past when you overflow your kernel
stack. The bottom of the kernel stack contains the threadinfo struct and
in that we have the task_cpu() data.

When we overflow the stack and overwrite the threadinfo we end up with 
a crazy number for task_cpu() and then oops when doing 
hardirq_ctx[task_cpu(p)]. Can you turn on DEBUG_STACKOVERFLOW and
DEBUG_STACK_USAGE and see if you get any stack overflow warnings? The
second option will also allow you to do sysrq t and dump the most stack
each process has used at that point.

Anton