Re: [PATCH] llist: Fix missing lockless_dereference()

[Mathieu Desnoyers]

----- Original Message -----
> From: "Michael Cree" <mcree@xxxxxxxxxxxx>
> To: "Mathieu Desnoyers" <mathieu.desnoyers@xxxxxxxxxxxx>
> Cc: "Greg KH" <gregkh@xxxxxxxxxxxxxxxxxxx>, linux-alpha@xxxxxxxxxxxxxxx, "Richard Henderson" <rth@xxxxxxxxxxx>, "Ivan
> Kokshaysky" <ink@xxxxxxxxxxxxxxxxxxxx>, "Matt Turner" <mattst88@xxxxxxxxx>, "Huang Ying" <ying.huang@xxxxxxxxx>,
> linux-kernel@xxxxxxxxxxxxxxx, "Paul McKenney" <paulmck@xxxxxxxxxxxxxxxxxx>, "David Howells" <dhowells@xxxxxxxxxx>,
> "Pranith Kumar" <bobby.prani@xxxxxxxxx>, stable@xxxxxxxxxxxxxxx
> Sent: Saturday, February 7, 2015 7:47:29 PM
> Subject: Re: [PATCH] llist: Fix missing lockless_dereference()
> 
> On Sat, Feb 07, 2015 at 10:30:44PM +0000, Mathieu Desnoyers wrote:
> > > On Fri, Feb 06, 2015 at 09:08:21PM -0500, Mathieu Desnoyers wrote:
> > > > A lockless_dereference() appears to be missing in llist_del_first().
> > > > It should only matter for Alpha in practice.
> 
> What could one anticipate to be the symptoms of such a missing
> lockless_dereference()?

This can trigger corruption of the lockless linked-list, which is
used across a few subsystems. AFAIU, the scenario is as follows.
Please bear with me, because it's been a while since I've read on
the Alpha multi-cache-banks behavior.

The list here would be initially non-empty. Initial state of
new_last->next is unset (newly allocated); IOW: garbage. CPU A
adds a node into the list while CPU B removes a node from the
head of the list.

CPU A                                      CPU B
llist_add_batch()
- Stores to new_last->next
- implicit full mb before cmpxchg makes the
  update to CPU A's cache bank containing
  new_last->next visible to other CPUs
  before CPU A's cache bank update making
  head->first visible to other CPUs.
- cmpxchg updates head->first = new_first
                                           llist_del_first()
                                           - entry = load head->first
                                           -> here, lack of barrier on Alpha creates a window where
                                              CPU B's cache bank can see the updated "head->first",
                                              but the cache bank holding the next value did not
                                              receive the update yet, since each cache bank have
                                              their own channel, which can be independently
                                              saturated.
                                           - next = load entry->next (dereference entry pointer)
                                           - cmpxchg updates head->first = next
                                             -> can store unset "next" value into head->first, thus
                                                corrupting the linked list.

> 
> The Alpha kernel is behaving pretty well provided one builds a machine
> specific kernel and UP.  When running an SMP kernel some packages
> (most notably the java runtime, but there are a few others) occasionally
> lock up in a pthread call --- could be a problem in libc rather then the
> kernel.

Are those lockups always occasional, or you have ways to reproduce them
frequently with stress-tests ?

Thanks,

Mathieu

> 
> > > Meta-comment, do we really care about Alpha anymore?  Is it still
> > > consered an "active" arch we support?
> 
> There are a few of us still running recent kernels on Alpha.  I am
> maintaining the unofficial Debian alpha port at debian-ports, and the
> Debian popcon shows about 10 installations of Debian Alpha.
> 
> Cheers
> Michael.
> 

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com
--
To unsubscribe from this list: send the line "unsubscribe linux-alpha" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html