Re: [RFC PATCH] LTTng instrumentation mm (updated)

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Dave Hansen (haveblue@us.ibm.com) wrote:
> On Tue, 2007-12-04 at 14:25 -0500, Mathieu Desnoyers wrote:
> > 
> > - I also dump the equivalent of /proc/swaps (with kernel internal
> >   information) at trace start to know what swap files are currently
> >   used.
> 
> What about just enhancing /proc/swaps so that this information can be
> useful to people other than those doing traces?
> 

It includes an in-kernel struct file pointer, exporting it to userspace
would be somewhat ugly.

> Now that we have /proc/$pid/pagemap, we expose some of the same
> information about which userspace virtual addresses are stored where and
> in which swapfile.  
> 

The problems with /proc :

- It exports all the data in formatted text. What I need for my traces
  is pure binary, compact representation.
- It's not very neat to export in-kernel pointer information like a
  kernel tracer would need.
- The locking is very often wrong. I started correcting /proc/modules a
  while ago, but I fear there are quite a few cases where a procfile
  reader could release the locks between two consecutive reads of the
  same list and therefore cause missing information or corruption. While
  being manageable for a proc text file, this is _highly_ unwanted in a
  trace. See my previous "seq file sorted" and "module.c sort module
  list" patches about this. My tracer deals with addition/removal of
  elements to a list between dumps done by "chunks" by tracing the
  modifications done to the list at the same time. However, /proc seq
  files will just get corrupted or forget about an element not touched
  by the modification, which my tracer cannot cope with.

Mathieu


> -- Dave
> 

-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68