On Thu, Aug 30, 2018 at 11:22:03AM -0700, Jason Thorpe wrote: > At the very least, an atime update should never drive an I/O.
Indeed. > Atime updates should update in-memory and *shrug*. If you want to > update the atime on stable storage when updating some other attribute, > so be it... and I suppose it wouldn't be too terrible to push out a > dirty atime when the vnode is reclaimed, but it's still unnecessary > write I/O for what might otherwise a read-only workload. I dimly remember part of my issues were with /dev on the noatime file system - so maybe we should just update atime in the vnode cache, but just don't write it back if mounted noatime. And maybe lock dev vnodes and never reclaim them - there can't be *that* much of them (if you have ptyfs mounted). Martin
