Created attachment 373753 [details]
Patch to fix the problem

Description of problem:

Previously, in RHEL5.3, there was just one [xenfb thread] in an RHEL5 Xen PV guest reported by ps and over subsequnent live migrations or save/restore cycles that one thread stayed.  Some time in RHEl5.4, I'm not sure when, the patch  linux-2.6-xen-fbfront-dirty-race.patch was introduced to fix bug 456893 (which I can't read) and this resulted in an initial two xenfb threads followed by another two every time the guest was live migrated or saved and restored.

Version-Release number of selected component (if applicable):
    kernel-xen-2.6.18-164.el5 and later (and possibly earlier)


How reproducible: Always


Steps to Reproduce:
1. ps -ef | grep xenfb
2. xm save; xm restore in dom0 or live migrate
3. ps -ef | grep xenfb
  
Actual results:
   At step 1 you'll see ps reporting two [xenfb thread] kernel threads, at step
   3 you'll see four of them.  If you repeat step 2 you'll get another two
   xenfb threads.   Eventually, I'm sure Bad Things Will Happen.

Expected results:
   Just one [xenfb thread]

Additional info:
   In linux-2.6-xen-fbfront-dirty-race.patch creation of the xenfb_thread
   thread is correctly deferred until XenbusStateConnected.  However, that
   state is entered every time the framebuffer is connected -- this seems to
   happen twice on initial boot and another twice after each successful
   save/restore or live migration.

   The attached patch simply avoids creating the thread if it already exists.
   There doesn't seem to be any possibility of introducing another race
   condition (the previous check for dev->state would seem to be enough of
   a barrier) and there also doesn't seem to be any need to create a new thread
   rather that using the existing one.