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

[Arnd Bergmann <arnd@arndb.de>]

On Sunday 28 March 2010, Frederic Weisbecker wrote:
> On Sun, Mar 28, 2010 at 09:05:50PM +0100, Arnd Bergmann wrote:
> > > 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',
> > which is almost identical, except for taking the inode mutex instead of the
> > BKL. 
> 
> 
> What if another file operation changes the file pointer while holding the bkl?
> You're not protected anymore in this case.
> 

Exactly, that's why I changed all the drivers to set default_llseek explicitly.
Even this is very likely not needed in more than a handful of drivers (if any),
for a number of reasons:

- sys_read/sys_write *never* hold any locks while calling file_pos_write(),
  which is the only place they get updated for regular files.
- concurrent llseek plus other file operations on the same file descriptor
  usually already have an undefined outcome.
- when I started inspecting drivers that look at file->f_pos themselves (not
  the read/write operation arguments), I found that practically all of them
  are doing this in a totally broken way!
- The only think we'd probably ever want to lock against in llseek
  is readdir, which is not used in any drivers, but only in file systems.

	Arnd