lkml.org 
[lkml]   [2008]   [Nov]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH -v3 7/8] fsnotify: add in inode fsnotify markings
From
Date
On Fri, 2008-11-28 at 05:42 +0000, Al Viro wrote:
> On Tue, Nov 25, 2008 at 12:21:28PM -0500, Eric Paris wrote:
>
> > +void fsnotify_mark_get(struct fsnotify_mark_entry *entry)
> > +{
> > + spin_lock(&entry->lock);
> > + entry->refcnt++;
> > + spin_unlock(&entry->lock);
> > +}
>
> > +void fsnotify_mark_put(struct fsnotify_mark_entry *entry)
> > +{
> > + spin_lock(&entry->lock);
> > + entry->refcnt--;
> > + /* if (!refcnt && killme) we are off both lists and nothing else can find us. */
> > + if ((!entry->refcnt) && (entry->killme)) {
> > + spin_unlock(&entry->lock);
> > + fsnotify_mark_kill(entry);
> > + return;
> > + }
> > + spin_unlock(&entry->lock);
> > +}
>
> Uh-huh... And what happens if fsnotify_mark_get() comes in the middle
> of final fsnotify_mark_put()? You spin on entry->lock, gain it just before
> fsnotify_mark_kill() which proceeds to kfree entry under you just as you
> increment its refcnt...

fsnotify_mark_get() can only find this object through either the
entry->i_list or entry->g_list. When we drop our ref to 0 and hold the
spinlock we know that no other task would have been able to find us on
those lists (everything that searches the i_list holds the
i_fsnotify_lock and that lock was dropped since we cleared ourselves
from that list and the same is true for the lock on the g_list side)

So if kill_me is set and the refcnt == 0 we are not on either list and
no other task could find this to try to call mark_get(). I'll review
it to make sure, but the design is that we are safe since nothing else
can find us to increment the ref cnt.
>
> > +void fsnotify_clear_mark_group(struct fsnotify_group *group)
> > +{
> > + struct fsnotify_mark_entry *entry;
> > + struct inode *inode;
> > +
> > + mutex_lock(&group->mark_mutex);
> > + while (!list_empty(&group->mark_entries)) {
> > + entry = list_first_entry(&group->mark_entries, struct fsnotify_mark_entry, g_list);
> > +
> > + /* make sure the entry survives until it is off both lists */
> > + fsnotify_mark_get(entry);
> > +
> > + /* remove from g_list */
> > + list_del_init(&entry->g_list);
> > + mutex_unlock(&group->mark_mutex);
> > +
> > + inode = entry->inode;
> > +
> > + spin_lock(&entry->lock);
> > + entry->killme = 1;
> > + spin_unlock(&entry->lock);
> >
> > + /* remove from i_list */
> > + spin_lock(&inode->i_fsnotify_lock);
>
> ... and just what would keep the inode from being freed under you here?

I'll review.



\
 
 \ /
  Last update: 2008-11-29 00:47    [W:0.196 / U:0.328 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site