Re: tests/misc/ls-time

[Jim Meyering <jim@meyering.net>]

Michael Stone wrote:
> It seems that touch -a does update ctime on btrfs, invalidating one of
> the assumptions behind this test and causing it to fail.

s/does/does not/

Thanks for the report.

I've just confirmed this test failure by building and running coreutils'
"make check" on a btrfs file system I'd just created using mkfs.btrfs
v0.19 on Fedora 12, btrfs-progs-0.19-9.fc12.x86_64

However, it has nothing to do with touch, but rather looks
like a bug (or at least a difference) in btrfs.

Here's a function to provide a quick demo:

ctime_vs_link_test() { env rm -f a b x; : > a; : > b; ln a x
stat -f --pr '%T: ' .; case "$(env ls -ct a b)" in a*b) echo pass;; \
  *) echo fail; env stat --format='%n %z' a b;; esac; }

Use it on a few file systems.
This test passes on all I tried except btrfs:
[note: $HOME is ext4]

    $ for i in $HOME/tmp /t /fs/btrfs /fs/xfs /fs/nilfs2; do
        cd $i && ctime_vs_link_test; done
    ext2/ext3: pass
    tmpfs: pass
    btrfs: fail
    a 2010-01-17 14:54:12.470194921 +0000
    b 2010-01-17 14:54:12.471193684 +0000
    xfs: pass
    nilfs2: pass

In case it's not immediately obvious (time-stamp problems rarely are),
here's a blow-by-blow:

    # After this, a's ctime precedes b's, even if just by ~1ms.
    rm -f a b x; : > a; : > b

    # Increasing a's link count must update its ctime to the present,
    # making it more recent than b's.
    ln a x

    # Sorting on ctime, a should come first:
    env ls -ct a b