[Ksummit-2008-discuss] DTrace

Roland Dreier rdreier at cisco.com
Fri Jun 27 14:12:07 PDT 2008


 > The basic problem here is that systemtap is adding debugging to kernel
 > functions ... if you don't know what the function is doing in a
 > reasonable amount of detail, it's not going to return much useful
 > information for you.

I think Solaris/DTrace is actually a counterargument to that: if the
people who *do* understand the kernel subsystem give you tracing
hooks/post-processing scripts, then those tracing hooks have a chance of
producing information at a level that a smart sysadmin can use.

And from the scripts quoted earlier:

 > dev_nr = $file->f_dentry->d_inode->i_sb->s_de

in addition to showing the level of sophistication required to write the
scripts, I think this proves Arjan's point that the tracing hooks have
to ship with the kernel -- because what are the chances of chains of
structure pointers like that remaining intact and keeping their meaning
between kernel versions?  Just think about the havoc that changes like
"remove struct class_device" would cause for out-of-tree tracing hooks.

 - R.


More information about the Ksummit-2008-discuss mailing list