Re: [0/3] Last 3 patches for bidi support

[Boaz Harrosh <bharrosh@panasas.com>]

On Tue, Nov 06 2007 at 20:25 +0200, Mike Christie <michaelc@cs.wisc.edu> wrote:
> Boaz Harrosh wrote:
>> [1]
>> I propose a small change to scsi_tgt_lib.c that will make
>> tgt completely neutral to the scsi_data_buffer patch. And will
>> make it all that more ready for bidi, too. TOMO is this OK?
>>
>> (Can you do without the GFP_KERNEL allocation flag? It could
>> make the code a bit more simple)
>>
> 
> GFP_KERNEL is nice for the target layer because it can sleep in that 
> path you changed and it and does not have the "cannot write out pages 
> because it may come back to the same device issues" like an initiator does.
> 
> If we ever changed to a softirq instead of the work queue then we would 
> not need the flag since it would have to GFP_ATOMIC, but I am not sure 
> if we have plans to do that anytime soon.

Yes I understand that, hence the GFP_KERNEL was kept intact in my patch.
But I was thinking perhaps it was possible to sleep outside, if
the return value was BLKPREP_DEFER, the way the block layer sleeps,
just not in allocation stage.

Boaz