Re: Question about the ide related ioctl's BLK* in 2.5.7-pre1 kernel

[Jens Axboe]

On Mon, Mar 18 2002, Martin Dalecki wrote:
> Jens Axboe wrote:
> >On Mon, Mar 18 2002, Martin Dalecki wrote:
> >[BLKRAGET etc]
> >
> >>BTW> It's quite propably right now, that I will just reintroduce them
> >>myself and give them the semantics of the multi-write hardware settings,
> >>just to fix the multi write PIO problem :-).
> >
> >
> >What would that fix?
> >
> >I've still got the multi-write fixes pending, out tomorrow I hope. Other
> >stuff keeps getting in the way. I haven't forgotten :-)
> 
> I think that it would make the write operation atomic in respect
> to the BIO chunk write order. Or please have a look at
> the following piece of cr... code from ide-taskfile.c:
> 
> 	/* (ks/hs): See task_mulin_intr */
> 	msect = drive->mult_count;
> 	nsect = rq->current_nr_sectors;
> 	if (nsect > msect)
> 		nsect = msect;
> 
> 	pBuf = ide_map_rq(rq, &flags);
> 	DTF("Multiwrite: %p, nsect: %d , rq->current_nr_sectors: %ld\n",
> 		pBuf, nsect, rq->current_nr_sectors);
> 	drive->io_32bit = 0;
> 	taskfile_output_data(drive, pBuf, nsect * SECTOR_WORDS);
> 	ide_unmap_rq(rq, pBuf, &flags);
> 	drive->io_32bit = io_32bit;
> 	
> 
> in esp. the nsect versus msect games whould go away and
> if we have current_nr_sectors > drive->mult_count, we
> are not going to write everything in one operation as it stands...
> The above just *feels* to me like something that should
> be pushed one layer upwards, since the sematics it has fits
> quite nicely into what the ioctl in question should be about.

You cannot guarentee that current_nr_sectors will be >= mult_count just
by doing that. One fix I did that was in the kernel for a while was to
just always setup the transfer for mult_count number of sectors at most.
This worked, but meant that we would potentially have to restart (output
command etc from scratch) a command a lot of times.

There a a lot of solutions to the problem, some better than others. The
whole issue boils dow to two things. One is that right after this
starts:

 	taskfile_output_data(drive, pBuf, nsect * SECTOR_WORDS);

you must be ready to handle the next transfer, ie the request state must
be sane. The other is that signalling request completion can get tricky,
as always with pio writes.

-- 
Jens Axboe