dd: introduce fast-stream skip

[oech3]

Assuming the situation downloading part of large file e.g.
curl * | dd skip=* bs=1M.

$ hyperfine "time head -c 1G /dev/zero |dd skip=1GiB bs=1M" "time head -c 1G /dev/zero |target/release/dd skip=1GiB bs=1M"
Benchmark 1: time head -c 1G /dev/zero |dd skip=1GiB bs=1M
  Time (mean ± σ):     407.0 ms ±  14.0 ms    [User: 10.6 ms, System: 788.1 ms]
  Range (min … max):   394.2 ms … 433.5 ms    10 runs
 
Benchmark 2: time head -c 1G /dev/zero |target/release/dd skip=1GiB bs=1M
  Time (mean ± σ):     235.7 ms ±   8.2 ms    [User: 3.2 ms, System: 451.5 ms]
  Range (min … max):   224.2 ms … 248.5 ms    13 runs
 
Summary
  time head -c 1G /dev/zero |target/release/dd skip=1GiB bs=1M ran
    1.73 ± 0.08 times faster than time head -c 1G /dev/zero |dd skip=1GiB bs=1M

[github-actions]

GNU testsuite comparison:

Skip an intermittent issue tests/misc/io-errors (fails in this run but passes in the 'main' branch)
Congrats! The gnu test tests/tail/tail-n0f is now passing!

[oech3]

GNU test failed: tests/dd/no-allocate.

[collinfunk]

This violates POSIX requirements:

skip=n
Skip n input blocks (using the specified input block size) before starting to copy. On seekable files, the implementation shall read the blocks or seek past them; on non-seekable files, the blocks shall be read and the data shall be discarded.

[oech3]

Cannot we consider splice length == block size?
[Edit] Actually, I was overreading (more than buf_size).

[oech3]

It seems tests/dd/no-allocate.sh does not allow skipping stream under ulimit with 0 allocation...

[oech3]

I don't think posix requires block should be existing on user space and read/write means read/write syscalls.
Also I think we can use sealed memfd for dd to exactly control block size. Rust std::io::copy's dev said splice to memfd is not zero-copy (thus, changing input during sinding data is safe), but still removes context switch.

Moreover, memfd is able to mmap and edit them for dd conv=.
@collinfunk ANy thought.