On Fri, Apr 25, 2025 at 06:10:14AM -0700, Christoph Hellwig wrote: > On Thu, Apr 24, 2025 at 03:00:41PM -0700, Darrick J. Wong wrote: > > The thing is, if you were relying on atime/mtime for detection of "file > > data changed since last read" then /not/ copying atime into the > > filesystem breaks that property in the image. > > I don't think that matter for images, because no software will keep > running over the upgrade of the image. Also plenty of people run > with noatime, and btrfs even defaulted to it for a while (not sure if > it still does). > > At the same time having the same behavior as mkfs.ext4 is a good thing > by itself because people obviously have been using it and consistency > is always a good thing. I don't see where mke2fs -d actually copies i_mtime into the filesystem. In misc/create_inode.c I see a lot of: now = fs->now ? fs->now : time(0); ext2fs_inode_xtime_set(&inode, i_atime, now); ext2fs_inode_xtime_set(&inode, i_ctime, now); ext2fs_inode_xtime_set(&inode, i_mtime, now); which implies that all three are set to a predetermined timestamp or the current timestamp. Also while I'm scanning create_inode.c, do you want to preserve hardlinks? > > How about copying [acm]time from the source file by default, but then > > add a new -p noatime option to skip the atime? > > I'd probably invert the polarity. When building an image keeping > atime especially and also ctime is usually not very useful. But that > would give folks who need it for some reason a way to do so. Either's fine with me. --D
