Here is a new perfmon patch. It is important because it
fixes the problem of the close() when the file descriptor
is shared between two related processes. The good thing
is that it simplifies a lot the cleanup of the sampling
buffer.

Here is the ChangeLog:

- fix bug in pfm_close() when the descriptor is
  shared between related processed. Introduce
  a pfm_flush() called for each invocation of
  close(). pfm_close() only called for the last
  user.

- fix pfm_restore_monitoring() to also reload
  the debug registers. They could be modified
  while monitoring is masked.

- fix pfm_close() to clear ctx_fl_is_sampling.

- fix a bug in pfm_handle_work() which could cause
  the wrong PMD to be reset. 

- converted PROTECT_CTX/UNPROTECT_CTX into
  local_irq_save/restore to keep context protection
  but allow IPI to proceed.

- updated pfm_syswide_force_stop() to use
  local_irq_save/restore now that the context
  is protected from the caller side.

- updated pfm_mck_pmc_check() to check if context is 
  loaded before checking for special IBR/DBR combinations.
  Clearing the debug registers is not needed when the context
  is not yet loaded.

- updated perfmon.h to have to correct prototype definitions
  for the pfm_mod_*() functions.

- got rid of the PFM_CTX_TERMINATED state.

- cleanup the DPRINT() statements to remove
  explicit output of current->pid. This is done
   systematically by the macros.

- added a systctl entry (expert_mode) to bypass
  read/write checks on PMC/PMD. As its name indicates
  this is for experts ONLY. Must be root to toggle
  /proc/sys entry.

- corrected pfm_mod_*() to check against the current task.

- removed pfm_mod_fast_read_pmds(). It is never needed.

- added pfm_mod_write_ibrs() and pfm_mod_write_dbrs().