Re: [PATCH 21/22] xfs: %pF is only for function pointers

[Scott Wood]

On Mon, 2015-08-31 at 18:06 +1000, Dave Chinner wrote:
> On Wed, Mar 11, 2015 at 10:13:56PM -0500, Scott Wood wrote:
> > Use %pS for actual addresses, otherwise you'll get bad output
> > on arches like ppc64 where %pF expects a function descriptor.
> > 
> > Signed-off-by: Scott Wood <scottwood@xxxxxxxxxxxxx>
> > Cc: xfs@xxxxxxxxxxx
> 
> Scott, I've just found that this change (commit 65dd297 "xfs: %pF is
> only for function pointers") breaks the symbolic printing in XFS
> trace events on x86_64. eg.
> 
> > diff --git a/fs/xfs/xfs_trace.h b/fs/xfs/xfs_trace.h
> > index 51372e3..b5ac81e 100644
> > --- a/fs/xfs/xfs_trace.h
> > +++ b/fs/xfs/xfs_trace.h
> > @@ -115,7 +115,7 @@ DECLARE_EVENT_CLASS(xfs_perag_class,
> >             __entry->refcount = refcount;
> >             __entry->caller_ip = caller_ip;
> >     ),
> > -   TP_printk("dev %d:%d agno %u refcount %d caller %pf",
> > +   TP_printk("dev %d:%d agno %u refcount %d caller %ps",
> >               MAJOR(__entry->dev), MINOR(__entry->dev),
> >               __entry->agno,
> >               __entry->refcount,
> 
> This results in output like this:
> 
> 760.828474: xfs_perag_get:  dev 253:32 agno 13 refcount 10 caller 
> 0xffffffff814eef02s
> 760.828476: xfs_perag_put:  dev 253:32 agno 13 refcount 9 caller 
> 0xffffffff814eefe8s
> 
> When I revert this commit, I get:
> 
> 71.911265: xfs_perag_get:   dev 253:32 agno 0 refcount 11 caller 
> xfs_extent_busy_insert
> 71.911266: xfs_perag_put:   dev 253:32 agno 0 refcount 10 caller 
> xfs_extent_busy_insert
> 
> Which is exactly what we should be getting from the tracing. I'm
> using trace-cmd to gather and print the events, and it breaks
> both old and current versions of trace-cmd.
> 
> Can you please look into why this change broke the tracing output
> on x86-64 - if there is no obvious/easy fix for it, then I'm simply
> going to revert it because having the tracing work correctly on
> x86-64 is far more important to us than ppc64 or ia64....

It looks like the cause is that TP_printk() is not really printk() -- it 
actually passes the format to userspace which has its own, not 100% 
compatible implementation pretty_print() in tools/lib/traceevent/event-
parse.c.  %pf in that function behaves like %ps in the kernel, and %ps is 
absent.

-Scott

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/