On Thu, Jan 28, 2021 at 1:57 PM Darrick J. Wong <djwong@xxxxxxxxxx> wrote: > > On Thu, Jan 28, 2021 at 08:46:04AM +0800, Nicolas Boichat wrote: [snip] > > Okay, so, based on this and Al's reply, I see 2 things we can do: > > 1. Go should probably not use copy_file_range in a common library > > function, as I don't see any easy way to detect this scenario > > currently (detect 0 size? sure, but that won't work with the example > > you provide above). And the man page should document this behaviour > > more explicitly to prevent further incorrect usage. > > 2. Can procfs/sysfs/debugfs and friends explicitly prevent usage of > > copy_file_range? (based on Al's reply, there seems to be no way to > > implement it correctly as seeking in such files will not work in case > > of short writes) > > One /could/ make those three provide a phony CFR implementation that > would return -EOPNOTSUPP, though like others have said, it's weird to > have regular files that aren't quite regular. Not sure where that > leaves them, though... Not that simple, as the issue happens on cross-filesystem operations where file_operations->copy_file_range is not called (and also, that'd require modifying operations for every single generated file... Anyway, made an attempt here: https://lore.kernel.org/linux-fsdevel/20210212044405.4120619-1-drinkcat@xxxxxxxxxxxx/T/#t > --D > > > > > Thanks, > > > > > > > > Cheers, > > > > > > Dave. > > > -- > > > Dave Chinner > > > david@xxxxxxxxxxxxx
