On Mon, 2023-09-18 at 11:24 -0700, Linus Torvalds wrote: > On Mon, 18 Sept 2023 at 04:54, Christian Brauner <brauner@xxxxxxxxxx> wrote: > > > > * Only update the atime if "now" is later than the current value. This > > can happen when the atime gets updated with a fine-grained timestamp > > and then later gets updated using a coarse-grained timestamp. > > I pulled this, and then I unpulled it again. > > I think this is fundamentally wrong. > > If somebody has set the time into the future (for whatever reason - > maybe the clocks were wrong at some point), afaik accessing a file > should reset it, and very much used to do that. > > Am I missing something? Because this really seems *horribly* broken garbage. > > Any "go from fine-grained back to coarse-grained" situation needs to > explicitly test *that* case. > > Not some kind of completely broken "don't update to past value" like this. > Fair point. Now that I've considered it some more, I think that commit 7df48e7d99a4 (fs: don't update the atime if existing atime is newer than "now") is not necessary. What prompted this was a bug report from the kernel test robot that showed the atime going backward on a STRICTATIME LTP test, but I think the root cause of that was the missing ctime initialization after allocation that we fixed in 0a22d3ff61b7 (fs: initialize inode->__i_ctime to the epoch). In general, we always update the atime with a coarse-grained timestamp, since atime and ctime updates are never done together during normal read and write operations. As you note, things are a little more murky with utimes() updates but I think we should be safe to overwrite the atime with a coarse-grained timestamp unconditionally. We should be able to just drop that patch from the series. Whether you want to pull the rest now or wait for a bit, I'll leave up to you and Christian to decide. Thanks, -- Jeff Layton <jlayton@xxxxxxxxxx>
