Re: + kernel-locking-lockdepc-make-lockdep-initialize-itself-on-demand.patch added to -mm tree

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 3 Feb 2016 08:44:30 +0100 Ingo Molnar <mingo@kernel.org> wrote:

> > Mike said:
> > 
> > : CONFIG_UBSAN_ALIGNMENT breaks x86-64 kernel with lockdep enabled, i.  e
> > : kernel with CONFIG_UBSAN_ALIGNMENT fails to load without even any error
> > : message.
> > : 
> > : The problem is that ubsan callbacks use spinlocks and might be called
> > : before lockdep is initialized.  Particularly this line in the
> > : reserve_ebda_region function causes problem:
> > : 
> > : lowmem = *(unsigned short *)__va(BIOS_LOWMEM_KILOBYTES);
> > : 
> > : If i put lockdep_init() before reserve_ebda_region call in
> > : x86_64_start_reservations kernel loads well.
> > 
> > Fix this ordering issue permanently: change lockdep so that it ensures
> > that the hash tables are initialized when they are about to be used.
> > 
> > The overhead will be pretty small: a test-n-branch in places where lockdep
> > is about to do a lot of work anyway.
> > 
> > Possibly lockdep_initialized should be made __read_mostly.
> > 
> > A better fix would be to simply initialize these (32768 entry) arrays of
> > empty list_heads at compile time, but I don't think there's a way of
> > teaching gcc to do this.
> > 
> > We could write a little script which, at compile time, emits a file
> > containing
> > 
> > 	[0] = LIST_HEAD_INIT(__chainhash_table[0]),
> > 	[1] = LIST_HEAD_INIT(__chainhash_table[1]),
> > 	...
> > 	[32767] = LIST_HEAD_INIT(__chainhash_table[32767]),
> > 
> > and then #include this file into lockdep.c.  Sounds like a lot of fuss.
> > 
> 
> ...
>
> Yuck, I don't really like this.
> 
> Lockdep initialization must happen early on,

It should happen at compile time.

> and it should happen in a well 
> defined place, not be opportunistic (and relatively random) like this, making it 
> dependent on config options and calling contexts.

That's an unusable assertion, sorry.

Initializing lockdep in the above manner guarantees that it is initialized
before it is used.  It is *much* more reliable than "try to initialize
it before some piece of code which hasn't even been written yet tries
to take a lock".

The conceptual problem is that if some piece of code does
spin_lock_init() or DEFINE_SPINLOCK(), that lock isn't necessarily
initialized yet.