Andrew,
This patch fixes queer behavior in __wait_on_freeing_inode().
If I_LOCK was not set it called yield(), effectively busy waiting for
the removal of the inode from the hash. This change was introduced
within "[PATCH] eliminate inode waitqueue hashtable" Changeset
1.1938.166.16 last october by wli.
The solution is to restore the old behavior, of unconditionally
waiting on the waitqueue. It doesn't matter if I_LOCK is not set
initally, the task will go to sleep, and wake up when wake_up_inode()
is called from generic_delete_inode() after removing the inode from
the hash chain.
Comment is also updated to better reflect current behavior.
Compile tested only. This condition is very hard to trigger normally
(simultaneous clear_inode() with iget()) so probably only heavy stress
testing can reveal any change of behavior.
Signed-off-by: Miklos Szeredi <[email protected]>
--- inode.c~ 2005-05-02 11:24:49.000000000 +0200
+++ inode.c 2005-05-06 13:25:12.000000000 +0200
@@ -1253,29 +1253,21 @@
}
/*
- * If we try to find an inode in the inode hash while it is being deleted, we
- * have to wait until the filesystem completes its deletion before reporting
- * that it isn't found. This is because iget will immediately call
- * ->read_inode, and we want to be sure that evidence of the deletion is found
- * by ->read_inode.
+ * If we try to find an inode in the inode hash while it is being
+ * deleted, we have to wait until the filesystem completes its
+ * deletion before reporting that it isn't found. This function waits
+ * until the deletion _might_ have completed. Callers are responsible
+ * to recheck inode state.
+ *
+ * It doesn't matter if I_LOCK is not set initially, a call to
+ * wake_up_inode() after removing from the hash list will DTRT.
+ *
* This is called with inode_lock held.
*/
static void __wait_on_freeing_inode(struct inode *inode)
{
wait_queue_head_t *wq;
DEFINE_WAIT_BIT(wait, &inode->i_state, __I_LOCK);
-
- /*
- * I_FREEING and I_CLEAR are cleared in process context under
- * inode_lock, so we have to give the tasks who would clear them
- * a chance to run and acquire inode_lock.
- */
- if (!(inode->i_state & I_LOCK)) {
- spin_unlock(&inode_lock);
- yield();
- spin_lock(&inode_lock);
- return;
- }
wq = bit_waitqueue(&inode->i_state, __I_LOCK);
prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
spin_unlock(&inode_lock);
? ???, 06/05/2005 ? 13:46 +0200, Miklos Szeredi ?????:
> This patch fixes queer behavior in __wait_on_freeing_inode().
Although the patch looks sane & simple, I gonna test your patch today.
I'll report the results.
--
Best Regards,
Artem B. Bityuckiy,
St.-Petersburg, Russia.
> > This patch fixes queer behavior in __wait_on_freeing_inode().
> Although the patch looks sane & simple, I gonna test your patch today.
> I'll report the results.
Thanks!
Miklos
On Fri, 2005-05-06 at 13:46 +0200, Miklos Szeredi wrote:
> The solution is to restore the old behavior, of unconditionally
> waiting on the waitqueue. It doesn't matter if I_LOCK is not set
> initally, the task will go to sleep, and wake up when wake_up_inode()
> is called from generic_delete_inode() after removing the inode from
> the hash chain.
That's all well and good if it's actually generic_delete_inode() which
removes the inode from the hash chain. But if it's prune_icache() which
does that, you don't get the wakeup.
Applying Artem's patch will fix that.
--
dwmw2
> That's all well and good if it's actually generic_delete_inode() which
> removes the inode from the hash chain. But if it's prune_icache() which
> does that, you don't get the wakeup.
>
> Applying Artem's patch will fix that.
I think it should work without Artem's patch too, since prune_icache()
removes the inode from the hash chain at the same time (under
inode_lock) as changing it's state to I_FREEING. So the pruned inode
will never be seen by iget().
Miklos
> I think it should work without Artem's patch too, since prune_icache()
> removes the inode from the hash chain at the same time (under
> inode_lock) as changing it's state to I_FREEING. So the pruned inode
> will never be seen by iget().
>
I suppose this doesn't mean that your patch fixes my problem (it mustn't
I believe) ?
--
Best Regards,
Artem B. Bityuckiy,
St.-Petersburg, Russia.
> > I think it should work without Artem's patch too, since prune_icache()
> > removes the inode from the hash chain at the same time (under
> > inode_lock) as changing it's state to I_FREEING. So the pruned inode
> > will never be seen by iget().
> >
> I suppose this doesn't mean that your patch fixes my problem (it mustn't
> I believe) ?
You are right. They are completely orthogonal, except the fact that
with your patch, the iget() clear_inode() collision will be more
frequent, and thus the yield() hack would trigger more often.
Miklos
On Fri, 2005-05-06 at 15:38 +0200, Miklos Szeredi wrote:
> I think it should work without Artem's patch too, since prune_icache()
> removes the inode from the hash chain at the same time (under
> inode_lock) as changing it's state to I_FREEING. So the pruned inode
> will never be seen by iget().
Doh. Yes, of course -- it had temporarily escaped my notice that the
whole _point_ of Artem's patch is that prune_icache() is currently
broken in precisely the way you describe.
--
dwmw2
> Compile tested only. This condition is very hard to trigger normally
> (simultaneous clear_inode() with iget()) so probably only heavy stress
> testing can reveal any change of behavior.
Well, my stress test works fine with your patch. I've tested it on JFFS2
FS which works on top of RAM-emulated flash. I tried it together with my
patch since otherwise the stress test crashes due to the race that my
patch fixes.
On vanilla linux-2.6.11.5 the stress test usually crashes in about 5
minutes, but linux-2.6.11.5 + the 2 patches (as well as + only one my
patch) it is stable for 2 hours already.
--
Best Regards,
Artem B. Bityuckiy,
St.-Petersburg, Russia.
On Fri, May 06, 2005 at 01:46:38PM +0200, Miklos Szeredi wrote:
> If I_LOCK was not set it called yield(), effectively busy waiting for
> the removal of the inode from the hash. This change was introduced
> within "[PATCH] eliminate inode waitqueue hashtable" Changeset
> 1.1938.166.16 last october by wli.
> The solution is to restore the old behavior, of unconditionally
> waiting on the waitqueue. It doesn't matter if I_LOCK is not set
> initally, the task will go to sleep, and wake up when wake_up_inode()
> is called from generic_delete_inode() after removing the inode from
> the hash chain.
I was trying to preserve some (possibly misinterpreted) behavior but I
can't remember what anymore. Anyway, since it's misbehaving, nuke it.
-- wli