TAKE 988255 - fix instant oops with tracing enabled

TAKE 988255 - fix instant oops with tracing enabled

[Lachlan McIlroy]

fix instant oops with tracing enabled

We can only read inode->i_count if the inode is actually there and not
a NULL pointer.  This was introduced in one of the recent sync patches.

Signed-off-by: Christoph Hellwig <hch@lst.de>

Date:  Tue Oct 14 12:17:19 EST 2008
Workarea:  redback.melbourne.sgi.com:/home/lachlan/isms/2.6.x-sync
Inspected by:  lachlan
Author:  lachlan

The following file(s) were checked into:
  longdrop.melbourne.sgi.com:/isms/linux/2.6.x-xfs-melb


Modid:  xfs-linux-melb:xfs-kern:32315a
fs/xfs/linux-2.6/xfs_vnode.c - 1.164 - changed
http://oss.sgi.com/cgi-bin/cvsweb.cgi/xfs-linux/linux-2.6/xfs_vnode.c.diff?r1=text&tr1=1.164&r2=text&tr2=1.163&f=h
	- fix instant oops with tracing enabled

Re: TAKE 988255 - fix instant oops with tracing enabled

[Dave Chinner]

On Tue, Oct 14, 2008 at 12:17:47PM +1100, Lachlan McIlroy wrote:
> fix instant oops with tracing enabled
> 
> We can only read inode->i_count if the inode is actually there and not
> a NULL pointer.  This was introduced in one of the recent sync patches.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>

BTW, this now means that other pending patches won't apply cleanly....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: TAKE 988255 - fix instant oops with tracing enabled

[Lachlan McIlroy]

Dave Chinner wrote:
> On Tue, Oct 14, 2008 at 12:17:47PM +1100, Lachlan McIlroy wrote:
>> fix instant oops with tracing enabled
>>
>> We can only read inode->i_count if the inode is actually there and not
>> a NULL pointer.  This was introduced in one of the recent sync patches.
>>
>> Signed-off-by: Christoph Hellwig <hch@lst.de>
> 
> BTW, this now means that other pending patches won't apply cleanly....
> 

Yeah this patch hunk didn't apply but was easy to hand merge.  It also
didn't compile because of a missing comma from the 4th last line.  Can
you please make sure each of your patch series compile on their own?

@@ -84,25 +84,12 @@ vn_ioerror(
 
 #ifdef	XFS_INODE_TRACE
 
-/*
- * Reference count of Linux inode if present, -1 if the xfs_inode
- * has no associated Linux inode.
- */
-static inline int xfs_icount(struct xfs_inode *ip)
-{
-	struct inode *inode = VFS_I(ip);
-
-	if (!inode)
-		return atomic_read(&inode->i_count);
-	return -1;
-}
-
 #define KTRACE_ENTER(ip, vk, s, line, ra)			\
 	ktrace_enter(	(ip)->i_trace,				\
 /*  0 */		(void *)(__psint_t)(vk),		\
 /*  1 */		(void *)(s),				\
 /*  2 */		(void *)(__psint_t) line,		\
-/*  3 */		(void *)(__psint_t)xfs_icount(ip),	\
+/*  3 */		(void *)(__psint_t)atomic_read(&VFS_I(ip)->i_count) \
 /*  4 */		(void *)(ra),				\
 /*  5 */		NULL,					\
 /*  6 */		(void *)(__psint_t)current_cpu(),	\

Re: TAKE 988255 - fix instant oops with tracing enabled

[Dave Chinner]

On Tue, Oct 14, 2008 at 03:33:06PM +1000, Lachlan McIlroy wrote:
> Dave Chinner wrote:
>> On Tue, Oct 14, 2008 at 12:17:47PM +1100, Lachlan McIlroy wrote:
>>> fix instant oops with tracing enabled
>>>
>>> We can only read inode->i_count if the inode is actually there and not
>>> a NULL pointer.  This was introduced in one of the recent sync patches.
>>>
>>> Signed-off-by: Christoph Hellwig <hch@lst.de>
>>
>> BTW, this now means that other pending patches won't apply cleanly....
>>
>
> Yeah this patch hunk didn't apply but was easy to hand merge.  It also
> didn't compile because of a missing comma from the 4th last line.  Can
> you please make sure each of your patch series compile on their own?

I do make sure every patch compiles given the config options a
mainline kernel allows. However, the CONFIG_XFS_TRACE option is not
present in mainline-derived kernels so it's not obvious when
problems occur inside code that isn't easily configurable on
such a kernel....

As it is, I haven't done a CONFIG_XFS_TRACE build for several months
because I've found that I haven't needed tracing to find problems
esoteric problems. Using watchpoints, breakpoints and single
stepping catches problems as they occur rather than needing tracing
to try to work out what went wrong after a problem has occurred.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: TAKE 988255 - fix instant oops with tracing enabled

[Timothy Shimmin]

Dave Chinner wrote:
> On Tue, Oct 14, 2008 at 03:33:06PM +1000, Lachlan McIlroy wrote:
>> Dave Chinner wrote:
>>> On Tue, Oct 14, 2008 at 12:17:47PM +1100, Lachlan McIlroy wrote:
>>>> fix instant oops with tracing enabled
>>>>
>>>> We can only read inode->i_count if the inode is actually there and not
>>>> a NULL pointer.  This was introduced in one of the recent sync patches.
>>>>
>>>> Signed-off-by: Christoph Hellwig <hch@lst.de>
>>> BTW, this now means that other pending patches won't apply cleanly....
>>>
>> Yeah this patch hunk didn't apply but was easy to hand merge.  It also
>> didn't compile because of a missing comma from the 4th last line.  Can
>> you please make sure each of your patch series compile on their own?
> 
> I do make sure every patch compiles given the config options a
> mainline kernel allows. However, the CONFIG_XFS_TRACE option is not
> present in mainline-derived kernels so it's not obvious when
> problems occur inside code that isn't easily configurable on
> such a kernel....
> 
> As it is, I haven't done a CONFIG_XFS_TRACE build for several months
> because I've found that I haven't needed tracing to find problems
> esoteric problems. Using watchpoints, breakpoints and single
> stepping catches problems as they occur rather than needing tracing
> to try to work out what went wrong after a problem has occurred.
> 
OOI, can anyone comment on any suggestions for other tracing mechanisms.
I recall Christoph mentioning something in the past but it wasn't
ready yet or something???.
I'd really like to be able to access tracing output via proc
instead of just within kdb,
as I have used in the past with Greg Banks qtrace patches.

--Tim

Re: TAKE 988255 - fix instant oops with tracing enabled

[Dave Chinner]

On Tue, Oct 14, 2008 at 05:12:08PM +1100, Timothy Shimmin wrote:
> Dave Chinner wrote:
> > On Tue, Oct 14, 2008 at 03:33:06PM +1000, Lachlan McIlroy wrote:
> >> Dave Chinner wrote:
> >>> On Tue, Oct 14, 2008 at 12:17:47PM +1100, Lachlan McIlroy wrote:
> >>>> fix instant oops with tracing enabled
> >>>>
> >>>> We can only read inode->i_count if the inode is actually there and not
> >>>> a NULL pointer.  This was introduced in one of the recent sync patches.
> >>>>
> >>>> Signed-off-by: Christoph Hellwig <hch@lst.de>
> >>> BTW, this now means that other pending patches won't apply cleanly....
> >>>
> >> Yeah this patch hunk didn't apply but was easy to hand merge.  It also
> >> didn't compile because of a missing comma from the 4th last line.  Can
> >> you please make sure each of your patch series compile on their own?
> > 
> > I do make sure every patch compiles given the config options a
> > mainline kernel allows. However, the CONFIG_XFS_TRACE option is not
> > present in mainline-derived kernels so it's not obvious when
> > problems occur inside code that isn't easily configurable on
> > such a kernel....
> > 
> > As it is, I haven't done a CONFIG_XFS_TRACE build for several months
> > because I've found that I haven't needed tracing to find problems
> > esoteric problems. Using watchpoints, breakpoints and single
> > stepping catches problems as they occur rather than needing tracing
> > to try to work out what went wrong after a problem has occurred.
> > 
> OOI, can anyone comment on any suggestions for other tracing mechanisms.
> I recall Christoph mentioning something in the past but it wasn't
> ready yet or something???.

Well, I'd suggest that we probably should turn all the trace points
into markers so that systemtap can hook them and you can write
probes to extract the traces in real-time....

> I'd really like to be able to access tracing output via proc
> instead of just within kdb,
> as I have used in the past with Greg Banks qtrace patches.

Perhaps extending something like:

Documentation/ftrace.txt

would be the appropriate way to go to acheive this.

I think in the past Christoph has suggested switching the ktrace
code over a relayfs mechanism and providing a userspace decoder
would allow both kdb and userspace to access the same traces...

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

Re: TAKE 988255 - fix instant oops with tracing enabled

[Christoph Hellwig]

On Tue, Oct 14, 2008 at 05:12:08PM +1100, Timothy Shimmin wrote:
> OOI, can anyone comment on any suggestions for other tracing mechanisms.
> I recall Christoph mentioning something in the past but it wasn't
> ready yet or something???.
> I'd really like to be able to access tracing output via proc
> instead of just within kdb,
> as I have used in the past with Greg Banks qtrace patches.

Yah, I've been working on doing that and wire it up with trace_mark.
I'll move it up on my TODO list a little..