lkml.org 
[lkml]   [2011]   [Nov]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 16/28] namespace: checkpatch wanking
On Tue, Nov 29, 2011 at 07:39:11PM -0800, Joe Perches wrote:

> No worries.
>
> I think patches 1 thru 14 are reasonable though
> and do apply with a few offsets to vfsmount-guts.

OK... BTW, speaking of this one:

-static int mnt_id_start = 0;
+static int mnt_id_start;
static int mnt_group_start = 1;

this deserves a separate NAK; explicit init here is actually more clear -
the variable is more or less analogous to mnt_group_start and it's less
distracting to have initializers spelled out in similar way as well.

- if ((child_mnt = __lookup_mnt(path->mnt, path->dentry, 1)))
+ child_mnt = __lookup_mnt(path->mnt, path->dentry, 1);
+ if (child_mnt)
mntget(child_mnt);
br_read_unlock(&vfsmount_lock);
return child_mnt;
Maybe, maybe not; in the #vfsmount-guts I seriously pondered turning that
into
child_mnt = ...;
if (child_mnt) {
mnt_add_count(child_mnt);
br_read_unlock(...);
return &child_mnt->mnt;
} else {
br_read_unlock(...);
return NULL;
}
Hell knows... Since __lookup_mnt() returns struct mount now *and* lookup_mnt()
returns struct vfsmount (much shrunk subset of the original; almost nothing
in there is needed for anything outside of core VFS code) we need if or ?:
after unlock as well. So it might make sense to split the codepaths once...
- if (flags & MNT_FORCE && sb->s_op->umount_begin) {
+ if (flags & MNT_FORCE && sb->s_op->umount_begin)
sb->s_op->umount_begin(sb);
- }
Probably. It used to be just a method call, then it had grown lock/unlock
around it, then lock went down into the method instances and the compound
statement stayed.

-static int do_add_mount(struct vfsmount *newmnt, struct path *path, int mnt_flags)
+static int
+do_add_mount(struct vfsmount *newmnt, struct path *path, int mnt_flags)

I don't mind BSD style like that, but again, check #vfsmount-guts; this
one takes struct mount * now, so that line happens to be below 80 cols.

-static int select_submounts(struct vfsmount *parent, struct list_head *graveyard)
+static int
+select_submounts(struct vfsmount *parent, struct list_head *graveyard)

Ditto.

- struct vfsmount *mnt = list_entry(tmp, struct vfsmount, mnt_child);
+ struct vfsmount *mnt = list_entry(tmp, struct vfsmount,
+ mnt_child);

Ditto (mnt_child moved into struct mount now)

- if (!(page = __get_free_page(GFP_KERNEL)))
+ page = __get_free_page(GFP_KERNEL);
+ if (!page)
return -ENOMEM;
Not sure... In general I'd agree, but in this case...


\
 
 \ /
  Last update: 2011-11-30 07:53    [from the cache]
©2003-2011 Jasper Spaans