Because Debian coreutils 9.1-1 "cp" silently falls back to copy_file_range()
after FICLONE reports EIO, "cp" can transfer incorrect bytes. Linux syscalls
having a file descriptor parameter report EIO after a fault in the underlying
device. The affected file is not recoverable in the general case, but syscall
outcomes after the EIO don't reflect that. For example, consider FICLONE
returning EIO for a fault during source file writeback. The kernel will mark
"clean" the affected page cache entries and clear the EIO state. If the page
cache evicts those pages, their file offsets revert to the last written-back
values if any, else zeros. If userspace issues a syscall that bypasses the
page cache, like copy_file_range() or another FICLONE, that syscall clones the
last written-back state or zeroes. See
https://lore.kernel.org/linux-xfs/20221108172436.GA3613139@rfd.leadboat.com
for a "cp" and "cat" test script, background, and discussion.
I recommend instead reporting the EIO and terminating when FICLONE or
copy_file_range() fails with EIO. One could argue that ENOSPC also warrants
termination, since no fallback reduces space usage. For other errno values,
fallback to the next transfer strategy, like today. An alternative would be
to fallback from FICLONE to copy_file_range() only after known-appropriate
errors EBADF, EINVAL, EOPNOTSUPP, ETXTBSY, and EXDEV. That alternative wins
if future FICLONE reports an additional termination-deserving errno value.
Since just EIO needs termination today, I bet new errno values are more likely
than not to deserve fallback. What do you think?
copy-ficlone-fail.patch