为什么 Linux 命令 cp 不是零拷贝？

Question

I use strace to trace command cp a.txt b.txt .

It shows that it copied the file using read and write syscall.

But there are some zero-copy syscall like sendfile, splice. Although splice does not support file-to-file data copy.

Why does not the command cp using those syscall (so we don't have to copy data to userspace)?

Answer 1

You answered your own question: sendfile() does not support file-to-file operation.

Why does not the Linux kernel provide a syscall which performs file to file data copy with zero-copy?

Probably because nobody has a use case where it improves performance by that much or in a way that doesn't have bad trade-offs. All you can really do is have a single syscall arrange both the DMA from disk to RAM and from RAM to disk, instead of two syscalls as you have now. If this reduction is syscalls matters to you, look at io_uring: https://vorner.github.io/2019/11/03/io-uring-mental-experiments.html (a very new feature, so you can't expect to see it used in cp yet).

People who care a lot about file copy performance often use copy-on-write to avoid copying in the first place, and/or break their files into multiple pieces so they only need to copy or modify one piece instead of a huge file.

Answer 2

https://man7.org/linux/man-pages/man2/sendfile.2.html

In Linux kernels before 2.6.33, out_fd must refer to a socket. Since Linux 2.6.33 it can be any file . If it is a regular file, then sendfile() changes the file offset appropriately.