Re: kblockd/1: page allocation failure in 2.6.9

[James Bottomley <James.Bottomley@SteelEye.com>]

On Fri, 2004-12-24 at 14:20 +0100, Jens Axboe wrote:
> This looks fishy - this is GFP_ATOMIC | GFP_DMA, where it should only be
> GFP_ATOMIC. gdth should not have shost->cmd_pool->gfp_mask ==
> GFP_ATOMIC, that looks like a bug in the driver.

This may be a fault in the gdth driver (if that's where the trace came
from).  It sets unchecked_isa_dma to 1 in its template and then resets
it at various places in the driver it may have been reset too late to
prevent it from getting the GFP_DMA scsi command pool.

> Apart from that, the trace looks sane and the SCSI mid layer should
> recover from this condition and not cause a hung io subsystem. The only
> way I can see this fail is if the scsi host free_list is not filled for
> some reason during init, or if the commands allocated from there are
> lost or never finished by the hardware.

Yes, I've checked this out in SCSI by artificially simulating a memory
allocation failure in scsi_get_command().  My system behaves nicely even
as I rack up the failures.  Is there anything else unusual in the log
that may indicate what the problem is?

James