Re: OOM Killer killing whole system

[James Bottomley]

On Fri, 2006-01-20 at 16:50 -0800, Andrew Morton wrote:
> For linux-scsi reference, Chase's /proc/slabinfo says:
> 
> scsi_cmd_cache    1547440 1547440    384   10    1 : tunables   54   27    8 : 
> slabdata 154744 154744      0

There's another curiosity about this: the linux command stack is pretty
well counted per scsi device (it's how we control queue depth), so if a
driver leaks commands we see it not by this type of behaviour, but by
the system hanging (waiting for all the commands the mid-layer thinks
are outstanding to return).  So, the only way we could leak commands
like this is in the mid-layer command return logic ... and I can't find
anywhere this might happen.

The sequence is:

driver -> cmd->scsi_done() -> blk softirq -> scsi_softirq_done() ->
scsi_finish_cmd() (where the queue counts are decremented, so anything
after here could leak commands if the rest of the chain is broken) ->
cmd->done() (which is the ULD completion callback) ->
scsi_io_completion() (frees the sg table, so if the sgpool slabs aren't
out of whack we must be past here) -> scsi_end_request() ->
scsi_next_command() -> scsi_put_command() (which is where the command
goes back to the slab).

James


> > Curious - the -s... were you expecting the ring buffer 
> > to exceed 16384?
> 
> It can sometimes be quite large.  I always say -s 1000000 to make sure
> everything got there.
> 
> > I don't think my (boot time) buffer does.
> 
> It's compile-time configurable with CONFIG_LOG_BUF_SHIFT and boot-time
> configurable with log_buf_len=n.
> -
> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/