lkml.org 
[lkml]   [2015]   [Apr]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v5 05/10] seqlock: Better document raw_write_seqcount_latch()
On Tue, Apr 14, 2015 at 12:25:06PM +0200, Ingo Molnar wrote:
>
> * Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
>
> > On Mon, Apr 13, 2015 at 11:21:46AM -0700, Linus Torvalds wrote:
> > > On Mon, Apr 13, 2015 at 10:43 AM, Paul E. McKenney
> > > <paulmck@linux.vnet.ibm.com> wrote:
> > > >
> > > > A shorthand for READ_ONCE + smp_read_barrier_depends() is the shiny
> > > > new lockless_dereference()
> > >
> > > Related side note - I think people should get used to seeing
> > > "smp_load_acquire()". It has well-defined memory ordering properties
> > > and should generally perform well on most architectures. It's (much)
> > > stronger than lockless_dereference(), and together with
> > > smp_store_release() you can make rather clear guarantees about passing
> > > data locklessly from one CPU to another.
> > >
> > > I'd like to see us use more of the pattern of
> > >
> > > - one thread does:
> > >
> > > .. allocate/create some data
> > > smp_store_release() to "expose it"
> > >
> > > - another thread does:
> > >
> > > smp_load_acquire() to read index/pointer/flag/whatever
> > > .. use the data any damn way you want ..
> > >
> > > and we should probably aim to prefer that pattern over a lot of our
> > > traditional memory barriers.
> >
> > I couldn't agree more!
>
> /me too!
>
> > RCU made a similar move from open-coding smp_read_barrier_depends()
> > to using rcu_dereference() many years ago, and that change made RCU
> > code -much- easier to read and understand. I believe that moving
> > from smp_mb(), smp_rmb(), and smp_wmb() to smp_store_release() and
> > smp_load_acquire() will provide similar maintainability benefits.
> > Furthermore, when the current code uses smp_mb(),
> > smp_store_release() and smp_load_acquire() generate faster code on
> > most architectures.
>
> A similar maintainability argument can be made for locking:
> spin_lock(x) was a big step forward compared to lock_kernel(),
> primarily not because it improves scalability (it often does), but
> because the '(x)' very clearly documents the data structure that is
> being accessed and makes locking and data access bugs a lot more
> visible in the review phase already.
>
> I wish rcu_read_lock() had a data argument, for similar reasons - even
> if it just pointed to a pre-existing lock or an rcu head it never
> touches ;-)

Heh! Jack Slingwine and I had that argument back in 1993. I advocated
placing the update-side lock into the rcu_read_lock() equivalent, and
he responded by showing me a use cases were (1) there were no update-side
locks and (2) there were many update-side locks, and it was impossible
to select just one on the read side. ;-)

However, DYNIX/ptx did not have anything like rcu_dereference() or
list_for_each_entry_rcu(), which perhaps can be used in your example
below. (Hey, that was 20 years ago, when 50MB was a lot of main
memory. So we relied on compilers being quite dumb.)

> As an example I picked a random file out of the kernel that uses RCU:
> kernel/cpuset.c::validate_change():
>
> static int validate_change(struct cpuset *cur, struct cpuset *trial)
> {
> struct cgroup_subsys_state *css;
> struct cpuset *c, *par;
> int ret;
>
> rcu_read_lock();
>
> /* Each of our child cpusets must be a subset of us */
> ret = -EBUSY;
> cpuset_for_each_child(c, css, cur)
> if (!is_cpuset_subset(c, trial))
> goto out;
>
> /* Remaining checks don't apply to root cpuset */
> ret = 0;
> if (cur == &top_cpuset)
> goto out;
>
> par = parent_cs(cur);
>
> /* On legacy hiearchy, we must be a subset of our parent cpuset. */
> ret = -EACCES;
> if (!cgroup_on_dfl(cur->css.cgroup) && !is_cpuset_subset(trial, par))
> goto out;
>
> /*
> * If either I or some sibling (!= me) is exclusive, we can't
> * overlap
> */
> ret = -EINVAL;
> cpuset_for_each_child(c, css, par) {
> if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
> c != cur &&
> cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
> goto out;
> if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
> c != cur &&
> nodes_intersects(trial->mems_allowed, c->mems_allowed))
> goto out;
> }
>
> /*
> * Cpusets with tasks - existing or newly being attached - can't
> * be changed to have empty cpus_allowed or mems_allowed.
> */
> ret = -ENOSPC;
> if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
> if (!cpumask_empty(cur->cpus_allowed) &&
> cpumask_empty(trial->cpus_allowed))
> goto out;
> if (!nodes_empty(cur->mems_allowed) &&
> nodes_empty(trial->mems_allowed))
> goto out;
> }
>
> /*
> * We can't shrink if we won't have enough room for SCHED_DEADLINE
> * tasks.
> */
> ret = -EBUSY;
> if (is_cpu_exclusive(cur) &&
> !cpuset_cpumask_can_shrink(cur->cpus_allowed,
> trial->cpus_allowed))
> goto out;
>
> ret = 0;
> out:
> rcu_read_unlock();
> return ret;
> }
>
> So just from taking a glance at that function can you tell me what is
> being RCU protected here? I cannot, I can only guess that it must
> either be cpuset_for_each_child() or maybe the cpumasks or other
> internals.
>
> And if I search the file for call_rcu() it shows me nothing. Only if I
> know that cpusets are integrated with cgroups and I search
> kernel/cgroup.c for call_rcu(), do I find:
>
> call_rcu(&css->rcu_head, css_free_rcu_fn);
>
> aha!
>
> ... or if I drill down 3 levels into cpuset_for_each_child() ->
> css_for_each_child() -> css_next_child() do I see the RCU iteration.

And I have felt that reviewing pain as well.

But shouldn't these API members be tagged with "_rcu" to make that
more clear? Sort of like the difference between list_for_each_entry
and list_for_each_entry_rcu()?

> It would have been a lot clearer from the onset, if I had a hint
> syntactically:
>
> rcu_read_lock(&css->rcu_head);
> ...
> rcu_read_unlock(&css->rcu_head);

I cannot resist asking what you put there if the update side uses
synchronize_rcu()... A NULL pointer? A pointer to synchronize_rcu()?
Something else? And what do you do in the not-uncommon case where
multiple RCU chains are being traversed in the same RCU read-side
critical section? One approach would be to use varargs, I suppose.
Though with a hash table, list, or tree, you could have a -lot- of
->rcu_head structures to reference, and concurrent additions and deletions
mean that you wouldn't necessarily know which at rcu_read_lock() time.

> beyond the reviewer bonus I bet this would allow some extra debugging
> as well (only enabled in debug kernels):
>
> - for example to make sure we only access a field if _that field_ is
> RCU locked (reducing the chance of having the right locking for
> the wrong reason)

One possibility would be to mark each traversal of an RCU-protected
pointer. Currently, if a multilinked structure is inserted in one
shot, only the initial pointer to that structure needs to have
rcu_dereference(). Otherwise, it is hard to tell exactly how far
the RCU protection is to extend. (Been having too much fun with
this sort of thing in the standards committees...)

> - we could possibly also build lockdep dependencies out of such
> annotated RCU locking patterns.

Tell me more?

> - RCU aware list walking primitives could auto-check that this
> particular list is properly RCU locked.

For example, that a lock in the proper update class was held during
the corresponding update?

> This could be introduced gradually by using a different API name:
>
> rcu_lock(&css->rcu_head);
> ...
> rcu_unlock(&css->rcu_head);
>
> (the 'read' is implied in RCU locking anyway.)

Agreed, a new API would be needed for something like this.

> ... and if you think this approach has any merit, I volunteer the perf
> and sched subsystems as guinea pigs! :-)

And rcutorture, not that it counts for much.

> What do you think?

I think that I don't yet fully understand your proposal. ;-)

Thanx, Paul



\
 
 \ /
  Last update: 2015-04-14 15:41    [W:0.121 / U:0.108 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site