Re: [discuss] Re: [PATCH] fs/compat.c: rwsem instead of BKL around ioctl32_hash_table

[Roland Dreier <roland@topspin.com>]

    Andi> The BKL couldn't protect again removal of sleeping compat
    Andi> handlers anyways because the BKL is dropped during a
    Andi> schedule, and they all can sleep in user accesses.  During
    Andi> this scheduling window the module could be unloaded if its
    Andi> count was zero. But with the assumption above this cannot
    Andi> happen.

    Andi> So basically the locking there is not to protect against
    Andi> running handlers, just to ensure consistency during the list
    Andi> walking. The list isn't touched anymore after the compat
    Andi> handler runs, so the sleeping in there is no problem.

    Andi> RCU could be used as well to protect the list because there
    Andi> is no sleep involved.

OK, good point.  My logic was wrong when I considered RCU.  But now I
don't understand what you meant by "it cannot work" when you wrote
this in your previous email:

    Andi> If you wanted to fix it properly better make it use RCU -
    Andi> but it cannot work for the case of calling a compat handler.

Based on what you just wrote, it seems to me like RCU would actually
work fine.  Is this wrong?

Thanks,
  Roland