Re: [GIT, RFC] Killing the Big Kernel Lock

[Jan Blunck <jblunck@suse.de>]

On Sun, Mar 28, Arnd Bergmann wrote:

> On Sunday 28 March 2010, Stefan Richter wrote:
> > Arnd Bergmann wrote:
> > > Your patches look good, but it would be helpful to also set .llseek = no_llseek
> > > in the file operations, because that is much easier to grep for than
> > > only the nonseekable_open. While it's technically a NOP on the presence of
> > > nonseekable_open, it will help that I don't accidentally apply my patch on
> > > top of yours.
> > 
> > Sounds like a plan, but (a) if my .llseek = no_llseek and your .llseek =
> > default_llseek are not within diff context range, you (or whoever else
> > merges mine and yours) only get a compiler warning (Initializer entry
> > defined twice) rather than a merge conflict which couldn't be missed,
> > (b) there won't be a merge conflict in "BKL removal: mark remaining
> > users as 'depends on BKL'".  (c) While I don't mind adding more visual
> > clutter to ieee1394/*, I prefer terse coding in firewire/*.
> > 
> > How about I put my nonseekable_open additions into a release branch and
> > send you a pull request after a few days exposure in linux-next?  If you
> > do not plan to respin your patch queue soon or at all, I could even let
> > you pull a for-arnd branch with a semantically correct merge of yours
> > and mine.
> 
> I can probably remember this specific one now, but for other people
> doing the same on their subsystems, adding no_llseek may help reduce
> the need for coordination.
> 
> > General thoughts:
> > 
> > ".llseek = NULL," so far meant "do the Right Thing on lseek() and
> > friends, as far as the fs core can tell".  Shouldn't we keep it that
> > way?  It's as close to other ".method = NULL," as it can get, which
> > either mean "silently skip this method if it doesn't matter" (e.g.
> > .flush) or "fail attempts to use this method with a fitting errno" (e.g.
> > .write).
> 
> My series changes the default from 'default_llseek' to 'generic_file_llseek',

That is not that easy. generic_file_llseek() is testing against 'offset <
inode->i_sb->s_maxbytes'. This is not necessarily true when you think about
directories with random offset cookies. I know that seeking on directories is
stupid but don't blame me.

> which is almost identical, except for taking the inode mutex instead of the
> BKL. Another option that has been discussed before is to make no_llseek
> the default, but that might cause more serious problems wiht drivers that
> really require seeking.
> 
> Since using default_llseek can only ever make a difference if the driver
> actually uses the BKL in any other function, I could go through the
> patches again and revert those that do no use the BKL anywhere else.
> 
> > Of course, as we have already seen with infiniband, firewire, ieee1394,
> > .llseek = NULL is ambiguous in practice.  Does the driver really want to
> > use default_llseek, or should it rather use no_llseek and/or
> > nonseekable_open, or should it even implement a dummy_llseek() { return
> > 0; } which avoids the BKL but preserves ABI behaviour?  This needs to be
> > resolved for each and every case eventually, regardless of whether or
> > when your addition of .llseek = default_llseek enters mainline.
> 
> Yes, that also sounds like a good idea. I believe that Jan actually posted
> a patch to do that at some point.

Yes, it is in

http://git.infradead.org/users/jblunck/linux-2.6.git bkl/default-lseek

There are some other patches in that branch that are not upstream yet. Mind to
take them for your bkl-removal branch?

Cheers,
Jan