Re: 2.5.9 -- OOPS in IDE code (symbolic dump and boot log included)

[Jens Axboe <axboe@suse.de>]

On Thu, Apr 25 2002, Martin Dalecki wrote:
> Uz.ytkownik Jens Axboe napisa?:
> >On Wed, Apr 24 2002, Martin Dalecki wrote:
> 
> >>OK I assume that the oops happens inside the ide-scsi module.
> >>This will be fixed in one of the forthcomming patch sets.
> >
> >
> >Are you sure this isn't just due to ->special being set, and
> >ide_end_request() assuming it's an ar? From ide-cd, that is.
> 
> 
> Yes I know it's all the same. However unfortunately
> it's *not easy* to back out the ->special use from
> the drivers that do it. We have the following sutuation:
> 
> 1. Generic BIO code checking for ->special and deciding whatever
> it should trying to merge request or not.
> 
> 2. Gneric ATA code setting ->special for ata_request passing.
> 
> 3. CD-ROM ATAPI code using ->special for passing packet commands
> and failed commands.
> 
> 4. ide-scsi using it for the same purspose as CD-ROM
> 
> 5. ide-floppy not using it at all buf abusing the ->buffer member
>    for precisely the same purpose.
> 
> And unfortunately there is *no* easy solution for any of the
> above circumstances without breaking far too many things.

You don't _have_ to back out the ->special usage. As I mentioned, it was
always just a quick hack for ide-disk so I didn't have to change every
single driver out there.

There are two options, as I see it:

- Keep ata_request as an ide-disk speciality. This is pretty trivial
  even though other drivers use ->special, because the ata_ar_put()
  path simply needs to do

	struct ata_request *ar = rq->special;

	if (ar && drive->media == ide_disk)
		ata_ar_put(ar);

  and that is it.

- Make the ata_request the general means of passing down request in the
  IDE layer -- start by making hwgroup->rq into hwgroup->ar and _never_
  store ar in ->special (you don't have to, you will always just go from
  ar -> rq, which is of course ar->ar_rq). This is what I wanted to do.

> The conclusion simply is: unless the above issues are fixed
> the TCQ stuff has simply to be backed out again anbd live
> separately from the main code chain. :-(.

If you didn't persist on pushing half-done stuff to Linus all the time,
I would have had the time to implement this properly... Now you keep
doing hackish work-arounds to make things limp along. So please calm
down for a moment, sick back, and think about it. It's a heck of a lot
better than going full throttle with an axe.

-- 
Jens Axboe