Hi,
I haven't followed the status of write barrier patches recently, I am
wondering if it's still "necessary" (to avoid file system corruption) to
disable the write cache of a SCSI disk drive when the machine doesn't
have an uninterruptible power supply or if instead the file systems and
driver know how to use ordered tags. (Fujitsu MAP drive: 8 MB cache,
AIC7880 adapter, SuSE Linux 8.2 patched 2.4.20 kernel with ext3 and xfs)
TIA,
--
Matthias Andree
IMO, it isn't "necessary", but it is very desirable, and should be the
default, to disable write cache on SCSI disks for any system that is
concerned about reliability.
If reliability is less important than performance, and you have more
sequential writes than random writes, then you would get a performance
boost, and those (limited) cases might want to turn write cache on. Disk
benchmark apps are one example.
However, most environments have a lot more small random writes than
sequential writes, and so don't even see a performance improvement from
turning disk write cache on.
The OS type or version shouldn't affect this, in principle.
Andy
-----Original Message-----
From: Matthias Andree [mailto:[email protected]]
Sent: Thursday, June 12, 2003 5:25 AM
To: Linux-Kernel mailing list
Subject: SCSI Write Cache Enable in 2.4.20?
Hi,
I haven't followed the status of write barrier patches recently, I am
wondering if it's still "necessary" (to avoid file system corruption) to
disable the write cache of a SCSI disk drive when the machine doesn't
have an uninterruptible power supply or if instead the file systems and
driver know how to use ordered tags. (Fujitsu MAP drive: 8 MB cache,
AIC7880 adapter, SuSE Linux 8.2 patched 2.4.20 kernel with ext3 and xfs)
TIA,
--
Matthias Andree
On Fri, 13 Jun 2003, Cress, Andrew R wrote:
> IMO, it isn't "necessary", but it is very desirable, and should be the
> default, to disable write cache on SCSI disks for any system that is
> concerned about reliability.
The question was if Linux was up to using ordered tags between the
requests, which would allow to use the write cache safely.
Cress, Andrew R wrote:
> IMO, it isn't "necessary", but it is very desirable, and should be the
> default, to disable write cache on SCSI disks for any system that is
> concerned about reliability.
This sounds like a bug, either in an application, or in Linux kernel's
scsi disk implementation.
Data is only guaranteed to be written onto disk following an
fsync(2)-like operation in the application. And in turn, it is the
Linux kernel's responsibility to ensure that such a flush is propagated
all the down to the low-level driver, in my opinion. Sophisticated
hosts can have barriers, and "dumb" hosts can simply call the drive's
flush-cache / sync-cache command.
Jeff
On Thu, 19 Jun 2003, Jeff Garzik wrote:
> This sounds like a bug, either in an application, or in Linux kernel's
> scsi disk implementation.
>
> Data is only guaranteed to be written onto disk following an
> fsync(2)-like operation in the application. And in turn, it is the
> Linux kernel's responsibility to ensure that such a flush is propagated
> all the down to the low-level driver, in my opinion. Sophisticated
I think file systems also have certain write ordering requirements to
maintain on-disk consistency, these would also need to make sure the
order is correct. AFAICS, SuSE have patched the reiserfs in their
2.4.20 kernel (8.2) to use write barriers (however deep these are
anchored), but ext3 or xfs don't show related log entries at boot-up or
mount time.
Is this something that will be fixed in 2.6 or will 2.6 still require me
to turn off the write cache?
> hosts can have barriers, and "dumb" hosts can simply call the drive's
> flush-cache / sync-cache command.
--
Matthias Andree