Re: Write cache on SATA drives?

[Jens Axboe <axboe@suse.de>]

On Fri, Jul 01 2005, Matthias Andree wrote:
> Jens Axboe schrieb am 2005-07-01:
> 
> > On Fri, Jul 01 2005, Matthias Andree wrote:
> > > Jens Axboe <axboe@suse.de> writes:
> > > 
> > > > On Thu, Jun 30 2005, Edwards, Scott (GE Healthcare) wrote:
> > > >> Hello,
> > > >> 
> > > >> Can anyone tell me if the write cache on SATA drives are handled such
> > > >> that journaling file systems work correctly?  Back in the 2.6.5 -
> > > >> 2.6.7 time frame I had to disable the write cache to get ext3 to not
> > > >> trash things when the power was lost.  With kernel 2.6.10 would I
> > > >> still need to disable the write cache?
> > > >
> > > > With 2.6.12.x it should work, if you use ext3 or reiser and the
> > > > appropriate mount options (-o barrier=1 for ext3, barrier=flush for
> > > > reiserfs). 2.6.11 and earlier does not work on SATA.
> > > 
> > > I presume this means "libata" exclusively, no?
> > 
> > It means "SCSI without queueing" which I'm assuming is basically only
> > libata at this moment, at least if you count setups that have a non-zero
> > userbase :)
> > 
> > > Which is the oldest version where this works
> > > 
> > > 1. for SCSI (perhaps by adaptor)?
> > 
> > 2.6.12, with the above restriction. I hope to lift that for .13/14, with
> > Tejuns barrier updates.
> 
> The "SCSI without queueing" isn't clear to me. What system exactly is
> meant "without queueing"?
> 
> Does this mean SCSI with TCQ is safe?
> 
> Does this means SCSI is safe only when TCQ isn't used?
> 
> Does this apply to internals of the SCSI host adaptor driver?
> 
> I'm not acquainted with kernel/block I/O queueing internals.

SCSI without TCQ or with TCQ depth=1 should be safe, provided the driver
never reorders a command once it has received it.

> > > 2. for traditional IDE (such as VIA 82*, PIIX_*)?
> > 
> > In SUSE kernels, for many years. Since 2.6.7/8'ish in Linus' kernels.
> 
> I don't care for vendor kernels. Is listing 2.6.8 safe?

Check the changelogs, I can't remember exactly if it was 2.6.7 or 2.6.8
(or perhaps .6...).

-- 
Jens Axboe