Re: [Jfs-discussion] Can JFS handle disk's write cache

[Andi K. <ak...@su...>]

On Thu, Jan 10, 2002 at 09:43:39AM -0600, Dave Kleikamp wrote:
> In short, JFS cannot guarantee the filesystem integrity when a disk's 
> write cache is on.

In theory 2.5 has the beginning of infrastructure for dealing with this.
The new bio layer supports passing of "barrier" requests down to the 
block device driver. This is just a normal request that is never merged
and has a special flag set. The barrier bit can also be set in the 
bio that is used to initiate IO from higher layer. 
Such a barrier could flush the write cache. 

JFS could use it with each journal write and only flush other data 
depending on it after the barrier has finished. 

Problems: 
AFAIK no device driver supports this yet.  Adding it would likely 
not be very hard though if the hardware supports it.

I was told that many IDE disks ignore the appropiate IDE command so that
there is no simple way for the driver to force a flush to platter.  It
may be possible to do some testing and white/black listing of disk 
drives here, but of course it would be a lot of effort.

-Andi