Re: [RFC 1/2] page-flags: Make page lock operation atomic

[Peter Zijlstra]

On Tue, Feb 12, 2019 at 08:45:35AM +0100, Jan Kara wrote:
> On Mon 11-02-19 09:56:53, Matthew Wilcox wrote:
> > On Mon, Feb 11, 2019 at 06:48:46PM +0100, Jan Kara wrote:
> > > On Mon 11-02-19 13:59:24, Linux Upstream wrote:
> > > > > 
> > > > >> Signed-off-by: Chintan Pandya <chintan.pandya@xxxxxxxxxxx>
> > > > > 
> > > > > NAK.
> > > > > 
> > > > > This is bound to regress some stuff. Now agreed that using non-atomic
> > > > > ops is tricky, but many are in places where we 'know' there can't be
> > > > > concurrency.
> > > > > 
> > > > > If you can show any single one is wrong, we can fix that one, but we're
> > > > > not going to blanket remove all this just because.
> > > > 
> > > > Not quite familiar with below stack but from crash dump, found that this
> > > > was another stack running on some other CPU at the same time which also
> > > > updates page cache lru and manipulate locks.
> > > > 
> > > > [84415.344577] [20190123_21:27:50.786264]@1 preempt_count_add+0xdc/0x184
> > > > [84415.344588] [20190123_21:27:50.786276]@1 workingset_refault+0xdc/0x268
> > > > [84415.344600] [20190123_21:27:50.786288]@1 add_to_page_cache_lru+0x84/0x11c
> > > > [84415.344612] [20190123_21:27:50.786301]@1 ext4_mpage_readpages+0x178/0x714
> > > > [84415.344625] [20190123_21:27:50.786313]@1 ext4_readpages+0x50/0x60
> > > > [84415.344636] [20190123_21:27:50.786324]@1 
> > > > __do_page_cache_readahead+0x16c/0x280
> > > > [84415.344646] [20190123_21:27:50.786334]@1 filemap_fault+0x41c/0x588
> > > > [84415.344655] [20190123_21:27:50.786343]@1 ext4_filemap_fault+0x34/0x50
> > > > [84415.344664] [20190123_21:27:50.786353]@1 __do_fault+0x28/0x88
> > > > 
> > > > Not entirely sure if it's racing with the crashing stack or it's simply
> > > > overrides the the bit set by case 2 (mentioned in 0/2).
> > > 
> > > So this is interesting. Looking at __add_to_page_cache_locked() nothing
> > > seems to prevent __SetPageLocked(page) in add_to_page_cache_lru() to get
> > > reordered into __add_to_page_cache_locked() after page is actually added to
> > > the xarray. So that one particular instance might benefit from atomic
> > > SetPageLocked or a barrier somewhere between __SetPageLocked() and the
> > > actual addition of entry into the xarray.
> > 
> > There's a write barrier when you add something to the XArray, by virtue
> > of the call to rcu_assign_pointer().
> 
> OK, I've missed rcu_assign_pointer(). Thanks for correction... but...
> rcu_assign_pointer() is __smp_store_release(&p, v) and that on x86 seems to
> be:
> 
>         barrier();                                                      \
>         WRITE_ONCE(*p, v);                                              \
> 
> which seems to provide a compiler barrier but not an SMP barrier? So is x86
> store ordering strong enough to make writes appear in the right order? So far
> I didn't think so... What am I missing?

X86 is TSO, and that is strong enough for this. The only visible
reordering on TSO is due to the store-buffer, that is: writes may happen
after later reads.