Re: [RFC] race in request_module()

[Keith Owens <kaos@ocs.com.au>]

On Tue, 23 Apr 2002 14:09:21 -0400 (EDT), 
Alexander Viro <viro@math.psu.edu> wrote:
>
>
>On Tue, 23 Apr 2002, Keith Owens wrote:
>
>>   open /dev/foo
>>     request_module(foo)
>>       load foo, mark !MOD_USED_ONCE.
>
>Another process:
>	open /dev/foo
>		MOD_INC_USE_COUNT
>		mark MOD_USED_ONCE
>	close /dev/foo
>		MOD_DEC_USE_COUNT
>rmmod -a
>	kills module
>
>>     continue with open, MOD_INC_USE_COUNT(foo), mark MOD_USED_ONCE.
>			  module not loaded

You need two rmmod -a sweeps with no intervening activity on the module
before the module will be autocleaned.  See MOD_VISITED.  Unless the
first process hangs in the kernel for minutes, it will find the module
still present, use the module and cancel the rmmod -a status.

BTW, I do not disagree that module unloading in general is racy,
especially for ancillary data such as exception tables, unwind data,
timers, kernel threads etc.  Which is why Rusty and I plan to rewrite
module loading and unloading in 2.5.  I just do not see a race in the
area you are looking at.