Re: lock_kernel() usage and sync_*() functions

[torvalds@transmeta.com (Linus Torvalds)]

In article <Pine.LNX.4.05.10103211901070.705-100000@cosmic.nrg.org>,
Nigel Gamble  <nigel@nrg.org> wrote:
>
>Why is the kernel lock held around sync_supers() and sync_inodes() in
>sync_old_buffers() and fsync_dev(), but not in sync_dev()?  Is it just
>to serialize calls to these functions, or is there some other reason?

A lot of the FS locks need the kernel lock and are not SMP-safe on their
own.  Look at "lock_super()" for the worst offender (I think most of the
other ones have been converted to properly lock on SMP).

sync_inodes() _shouldn't_ need it. sync_supers() definitely does.

The fact that sync_dev() doesn't get the kernel lock looks worrisome. 
Of course, I don't think much of anything actually _uses_ "sync_dev()"
anyway (quick grep shows it up in revalidate, which gets the kernel lock
earlier)

But it might be a good idea to try to (a) remove the bkl from the
functions, and push it down into sync_supers() that definitely needs it
now (and remove it when it doesn't any more).

The long-term plan (ie 2.5.x) is to basically remove the bkl from all
the VFS interfaces. For 2.4.x, only the truly performance-critical stuff
was done (ie mainly name lookup and read/write page).

		Linus