James Youngman <j...@gnu.org> wrote:
> [ adding nug-gnulib to CC since we're discussinf its fts implementation ]
>
> On Sun, Mar 8, 2009 at 2:50 PM, George Spelvin <li...@horizon.com> wrote:
>> James Youngman <invalid.nore...@gnu.org> wrote:
>>> The updated patches are what I actually pushed (into 4.5.x).
>>
>> Ah, thank you. The posted patches wouldn't apply to savannah git HEAD
>> due to a conflict in the ChangeLog. (Easy enough to work around, but
>> worrisome.)
>
> They're *in* HEAD.
Um, they weren't earlier today. (I cloned at 2009-03-08 11:49:25 UTC.)
>
> >
> > The (tested) simplified version is, of course,
> > + ?boolean todo = pred_ptr->need_stat ||
> > + ? ? ? ? ? ? ?(pred_ptr->need_type && !state.have_type) ||
> > + ? ? ? ? ? ? ?(pred_ptr->need_inum &&
> > + ? ? ? ? ? ? ? (!p->st_ino || !state.have_type || S_ISDIR(p->st_mode)));
> >
> > I poked at the leaf optimization, but it appears I'd have to change the
> > prototypes for the entire predicate system, which is extremely annoying.
>
> Yes. I also believe it would gain little, since in the case of
> ftsfind at least, leaf is normally 0.
Oh. Is that because ftsfind doesn't implement the optimization, or is there
some more fundamental reason?
>> One option is to understand the "state" structure and how its contents
>> relate to the predicate arguments. I'm currently very confused by that.
>
> The state structure contains a bunch of what would otherwise be global
> variables. The struct predicate instances represent tests or actions
> specified on the command line. With a few exceptions (for example
> struct predicate_performance_info) the information in struct predicate
> is unchanged after we begin traversing the filesystem.
Thank you for the explanation, but due to my sloppy phrasing, it's not to
what I was wondering about. I understand that the "struct predicate"
tree is constant; I was referring to the variety of
boolean
pred_FOO (const char *pathname, struct stat *stat_buf, struct predicate
*pred_ptr)
{
...
}
helper functions and those function arguments. Consider, for example, the
pred_type function, which begins:
boolean
pred_type (const char *pathname, struct stat *stat_buf, struct predicate
*pred_ptr)
{
mode_t mode;
mode_t type = pred_ptr->args.type;
assert (state.have_type);
if (0 == state.type)
{
/* This can sometimes happen with broken NFS servers.
* See Savannah bug #16378.
*/
return false;
}
(void) pathname;
if (state.have_stat)
mode = stat_buf->st_mode;
else
mode = state.type;
Notice how it seems to assume that state.have_stat describes the same
"current file" as the stat_buf argument.
So perhaps the leaf flag could be stuffed into the state structure.
(Ideally, the pathname and the stat_buf would be stuffed into the
state structure and a pointer to it would be passed down, so the
find engine would be reentrant.)
> FWIW, some broken NFS servers (or perhaps NFS servers serving
> corrupted filesystems) return invalid st_mode values; Savannah bug
> #16378 at https://savannah.gnu.org/bugs/?16378 gives an example.
Joy.
