[PATCH v2 00/15] Make the user mode driver code a better citizen

[Eric W. Biederman]

Just to make certain I understand what is going on I instrumented a
kernel with some print statements.

a) The workqueues and timers start before populate_rootfs.

b) populate_rootfs does indeed happen long before the bpfilter
   module is intialized.

c) What prevents populate_rootfs and the umd_load_blob from
   having problems when they call flush_delayed_put is the
   fact that fput_many does:
   "schedule_delayed_work(&delayed_fput_work,1)".

   That 1 requests a delay of at least 1 jiffy.  A jiffy is between
   1ms and 10ms depending on how Linux is configured.

   In my test configuration running a kernel in kvm printing to a serial
   console I measured 0.8ms between the fput in blob_to_mnt and
   flush_delayed_fput which immediately follows it.

   So unless the fput becomes incredibly slow there is nothing to worry
   about in blob_to_mnt.

d) As the same mechanism is used by populate_rootfs.  A but in the
   mechanism applies to both.

e) No one appears to have reported a problem executing files out of
   initramfs these last several years since the flush_delayed_fput was
   introduced.
 
f) The code works for me.  There is real reason to believe the code will
   work for everyone else, as the exact same logic is used by initramfs.
   So it should be perfectly fine for the patchset and the
   usermode_driver code to go ahead as written.

h) If there is something to be fixed it is flush_delayed_fput as that is
   much more important than anything in the usermode driver code.

Eric

p.s.) When I talked of restarts of the usermode driver code ealier I was
   referring to the code that restarts the usermode driver if it is
   killed, the next time the kernel tries to talk to it.

   That could mask an -ETXTBUSY except if it happens on the first exec
   the net/bfilter/bpfilter_kern.c:load_umh() will return an error.