On 09/12/15 10:35, Pádraig Brady wrote:
> On 09/12/15 06:34, Kamil Dudka wrote:
>> The flag is needed to implement the -noleaf option of find.
>> * lib/fts.c (link_count_optimize_ok): Implement the FTS_NOLEAF flag.
>> * lib/fts_.h (FTS_NOLEAF): New macro, shifted conflicting constants.
>
> Is this exposed to fix issues with certain file systems,
> or just in case there may be issues, or support easily
> testing find without the leaf optimization?
>
> I see Jim said the current FTS implementation
> would make -noleaf a no-op there:
> https://lists.gnu.org/archive/html/bug-gnulib/2008-12/msg00280.html
>
> cheers,
> Pádraig.
>
> p.s. I see that find does a stat per file on XFS,
> while d_type can be used to distinguish dirs there.
> On XFS DT_DIR is set for dirs and DT_UNKNOWN otherwise.
> I wonder is there some optimization we could do for that case.
I did a quick check on XFS which suggests the leaf optimization
based on st_nlink is valid:
test $(($(find . -maxdepth 1 -type d | wc -l) + 1)) = $(stat -c %h .) && echo
leaf_ok
Applying this diff:
@@ -717,6 +718,7 @@ leaf_optimization_applies (int dir_fd)
{
/* List here the file system types that lack usable dirent.d_type
info, yet for which the optimization does apply. */
+ case S_MAGIC_XFS:
case S_MAGIC_REISERFS:
return true;
Gives this significant speedup:
$ time find/find-before /usr/share >/dev/null
real 0m0.410s
user 0m0.145s
sys 0m0.266s
$ time find/find-after /usr/share >/dev/null
real 0m0.278s
user 0m0.147s
sys 0m0.131s
I also noticed a lot of fcntl calls on XFS
(basically one per file), which I need to look further into:
$ strace -c find/find /usr/share >/dev/null
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
40.03 0.147809 1 151710 fcntl
17.62 0.065069 1 63154 close
14.52 0.053608 1 40071 newfstatat
13.15 0.048547 1 35400 getdents
cheers,
Pádraig
