svn commit: r330436 - head/bin/chflags

[Bruce Evans]

On Sun, 4 Mar 2018, Ian Lepore wrote:

> On Mon, 2018-03-05 at 01:56 +0000, Bryan Drewery wrote:
>>
>> Log:
>> \xa0 chflags: Add -x option to not traverse mount points.
>
> Yay! \xa0One day later than I needed it, but still, yay!

I recently noticed that find(1) needs an option to not look at mount
points at all, and further options to classify mount points so that
you can prune them.

After reading the above and investigating further, I noticed that -x
is broken in most FreeBSD utilities, since POSIX requires not looking
at mount points at all for the few utilities that support -x.  E.g.,
for du in 2001 draft 7 POSIX:

12578            -x            When evaluating file sizes, evaluate only those files that have the same device as
12579                          the file specified by the file operand.
12580            Specifying more than one of the mutually-exclusive options -H and -L shall not be considered

Mount points are on a different device, but FreeBSD du -x is broken by
using FTS_XDEV and not filtering out the mount points.  Linux du -x
(an old version in /compat/linyx/usr/bin) works correctly.

(Elsewhere, POSIX uses the terms "mount point" and "file system" without
specifying what they mean.  E.g., for FTW_MOUNT which corresponds tp
FTS_XDEV, it says "The walk does not cross a mount point".  It is also
unclear what "cross" means here -- the names of mount points are in the
same file system although everything (?) else is in another file system.
"mount point" should be specified as "a file whose device differs from
the device of its directory", and "file system" should be specified as
"all files on the system with the same given device".  The latter allows
the trees for file system to have weird topologies but is as good as
anything.  It is better than "all files under a mount point with the
same device as the mount point".  With the latter definition, given
stat structs or even pathnames to 2 files, you would potentially have
to walk all files starting at the root directory (if there is one) to
determine if the 2 files satisfy the topological condtion of being
under the same mount point.)

FTS_XDEV is documented as doing the wrong thing for du -x by saying that
it doesn't "descend" "into directories that have a different device number
than the file from which the descent began".  Its description of
"same device" is too verbose.  It would benefit from formally specifying
mount points and file systems and "cross" too.  Its "descent" is just
good enough without a formal specification since the mount point is at
level 0 in the new file system and a descent of 0 levels is not a
descent in informal use (like for "negative", but unlike for ascent/positive
where it is unclear if 0 levels is an degenerate ascent).

The old POSIX draft uses "same.*device" just once according to a simple
grep.  This is the one quoted above for du -x.  Old POSIX find doesn't
even have -x.  It has -xdev.  This is strange, since FreeBSD's man page
has said since long before 2001 that it is -xdev that is deprecated.  It
specifies -xdev using "descending", like for FTS_XDEV but with worse verbose
wording.

find(1) seems to be missing a classification option for mount points in
both FreeBSD and POSIX.  This could be used to prune mount points, but
pruning is much harder than using a flag that works like -x should work.

Perhaps something can be done with pruning levels for any utility that
supports -x).  E.g. -xx<number> could descend to N levels, where the
level counter is set to 0 when a mount point is _first_ reached.  Then
-xx-1 would give the correct du -x behaviour of not doing the degenerate
descent to level 0; -xx0 would give the find behaviour of doing the
degenerate descent to level 0; -xx1 would descend 1 level; other
positive descents are easy to implement; ther negative descents are
too hard to implement.

Bruce