On Mon, Feb 23, 2009 at 09:23:56PM +0000, Andrew Doran wrote: > I didn't phrase that very well. I mean if you only read a file but never > dirty it, the atime update won't go out until unmount. Since writing the > original mail I learned that sync() won't even push it. This is now fixed. > The same issue also affected device nodes. This was deliberate; at the time it was a way to avoid spinning up disks on laptops, just for atime updates. The fact that sync didn't push them either was an unfortunate implementation issue. Yes, there should be A Better Way, allowing writes to be prioritised and classified such that they can trickle out while the disk is spinning, but not cause it to spin up until a higher-priority IO request comes. We have the beginnings of this now, but the fs consequences of holding those writes at the disk layer are unclear. -- dan.
Description: PGP signature