Re: Patch added to scsi-pending-2.6: [SCSI] scsi_dh: Use SCSI device handler in dm-multipath

[Chandra Seetharaman <sekharan@us.ibm.com>]

Hi Alasdair,

Thanks for your feedback.

MikeC and mikeand, feel free to add.

On Tue, 2008-04-29 at 15:56 +0100, Alasdair G Kergon wrote:
> On Thu, Apr 24, 2008 at 04:03:11PM +0000, James Bottomley wrote:
> >     This patch converts dm-mpath to scsi hw handlers. It does
> >     not add any new functionality and old behaviors and userspace
> >     tools work as is except we use the safe clariion default instead
> >     of using the userspace setting.
> 
> OK.  Comments from my first line-by-line reading of this.
> 
> Firstly, the patch header is rather too brief for my liking - the patch
> does make some subtle functional changes that ought to be commented upon
> to explain to people why they are being made.  (I would have preferred

will do.
> this patch to have been broken up into smaller ones in a manner that
> made the functional changes more obvious i.e. separated from the code
> reorganisation.)

will do.

> 
> 1. Currently the supplied hw_handler name is validated at table load
> time so userspace cannot create a device using a non-existent
> hw_handler.  After this patch it looks as if that validation is delayed
> until something starts to access the device - and then the paths simply
> get failed, a new failure mode for the userspace tools to handle.  Is
> that change hard to avoid or was it judged not worth the effort to leave
> things as they were and perform as much validation / resource allocation
> as possible up front?

For the current implementation, initialization of the scsi_dh specific
data need to happen when we see the device the very first time (for
prep_fn() and check_sense() to be useful). We cannot wait till multipath
comes along, and the device specific data will exist till the device is
removed, hence there is no need to hold onto a reference for each device
for mulipath's sake.

But, I agree with your point that the failure should happen early.

Will add a scsi_dh_handler_exist(name) function to make sure that the
module is available at the time of table load. What do you think ?

> 
> 2.  parse_hw_handler() now ignores hw_handler arguments.
> Please add a comment to explain that or perhaps log a warning message if
> any are supplied.

will do

> BTW r looks superfluous now:
> 
>         if (read_param(_params, shift(as), &hw_argc, &ti->error))
>                 return -EINVAL;

will fix.

> 
> 3. The dmsetup table output has changed - it no longer matches the input
> in the case where arguments got ignored.  I don't think this matters
> (unlike a similar issue we had with crypt), but it should be noted in
> comments inline.

will add comments.

> 
> 4.  /* Fields used by hardware handler usage */
> Comment doesn't add anything - drop it?
> 
ok.

> 5. char *hw_handler_name;
> const?

will fix.

> 
> 6. The code now assumes all hw_handlers have a pg_init function: in
> practice this is likely to be the case.  Does that permit further
> simplification of some of the logic?

will look into it.

> 
> 7. I note that the 'bypassed' argument to the pg_init function has been
> dropped - presumably as none of the existing handlers made any use of
> it.

Yes, none of the hardware handlers used it.

> 
> 8. A new workqueue is introduced without comment.

Will add.

>   Which of the changes means we need it now? 
I do not understand this comment.

>  What's the reason it's not single-threaded (and per-device)?
Per device might be an overkill. Will do single-threaded.

>   Is there a clearer name than "khwhandlerd" -
> something that tells people it's connected with multipath, and

how does kmpath_handlerd sound ?

> can the name of the struct workqueue_struct variable match this?

will fix.

> Previously the hw_handler pg_init function was required to return
> immediately.  Can its replacement scsi_dh_activate() block (but not in a
> way that could deadlock of course)? 

Yes, scsi_dh_activate() can block. But, at the dm-level, it is
indifferent, due to the usage of the workqueue.

>  Should some of its functionality be
> got out of the way at initialisation time within multipath_ctr()?

You mean hardware handler specific initialization ? It happens when the
device is found. We do not want to wait till multipath comes along.

> 
> 9. What does m->path_to_activate give us that m->current_pgpath
> doesn't?

I guess none. Was just sticking with the original interface as it was
handed off to an asynchronous thread. If it won't matter, I will fix it
to use it directly from the multipath data structure.

> 
> 10. Is activate_passive_path() an accurate function name?  Is the
> supplied path argument always 'passive'?

May be not always. Will change it activate_path().

> 
> 11. Where has the use of pg_init_retries gone?

Oops.. will fix.
> 
> I still need to check the updated state machine logic and associated
> locking once we have a final version of this patch.
> 
> Alasdair