Re: [RFC PATCH v2 3/4] hp: Implement Hazard Pointers

[Paul E. McKenney]

On Mon, Oct 07, 2024 at 10:50:46AM -0400, Mathieu Desnoyers wrote:
> On 2024-10-07 12:40, Peter Zijlstra wrote:
> > On Sat, Oct 05, 2024 at 02:50:17PM -0400, Mathieu Desnoyers wrote:
> > > On 2024-10-05 18:04, Peter Zijlstra wrote:
> > 
> > 
> > > > > +/*
> > > > > + * hp_allocate: Allocate a hazard pointer.
> > > > > + *
> > > > > + * Allocate a hazard pointer slot for @addr. The object existence should
> > > > > + * be guaranteed by the caller. Expects to be called from preempt
> > > > > + * disable context.
> > > > > + *
> > > > > + * Returns a hazard pointer context.
> > > > 
> > > > So you made the WTF'o'meter crack, this here function does not allocate
> > > > nothing. Naming is bad. At best this is something like
> > > > try-set-hazard-pointer or somesuch.
> > > 
> > > I went with the naming from the 2004 paper from Maged Michael, but I
> > > agree it could be clearer.
> > > 
> > > I'm tempted to go for "hp_try_post()" and "hp_remove()", basically
> > > "posting" the intent to use a pointer (as in on a metaphorical billboard),
> > > and removing it when it's done.
> > 
> > For RCU we've taken to using the word: 'publish', no?
> 
> I'm so glad you suggest this, because it turns out that from all
> the possible words you could choose from, 'publish' is probably the
> most actively confusing. I'll explain.
> 
> Let me first do a 10'000 feet comparison of RCU vs Hazard Pointers
> through a simple example:
> 
> [ Note: I've renamed the HP dereference try_post to HP load try_post
>   based on further discussion below. ]
> 
> *** RCU ***
> 
> * Dereference RCU-protected pointer:
>     rcu_read_lock();          // [ Begin read transaction ]
>     l_p = rcu_dereference(p); // [ Load p: @addr or NULL ]
>     if (l_p)
>       [ use *l_p ...]
>     rcu_read_unlock();        // [ End read transaction ]
> 
> * Publish @addr:    addr = kmalloc();
>                     init(addr);
>                     rcu_assign_pointer(p, addr);
> 
> * Reclaim @addr:    rcu_assign_pointer(p, NULL); // [ Unpublish @addr ]
>                     synchronize_rcu();           // Wait for all pre-existing
>                                                  // read transactions to complete.
>                     kfree(addr);
> 
> 
> *** Hazard Pointers ***
> 
> * Load and post a HP-protected pointer:
>     l_p = hp_load_try_post(domain, &p, &slot);
>     if (l_p) {
>       [ use *l_p ...]
>       hp_remove(&slot, l_p);
>     }
> 
> * Publish @addr:    addr = kmalloc();
>                     init(addr);
>                     rcu_assign_pointer(p, addr);
> 
> * Reclaim @addr:    rcu_assign_pointer(p, NULL); // [ Unpublish @addr ]
>                     hp_scan(domain, addr, NULL);
>                     kfree(addr);
> 
> Both HP and RCU have publication guarantees, which can in fact be
> implemented in the same way (e.g. rcu_assign_pointer paired with
> something that respects address dependencies ordering). A stronger
> implementation of this would be pairing a store-release with a
> load-acquire: it works, but it would add needless overhead on
> weakly-ordered CPUs.
> 
> How the two mechanisms differ is in how they track when it is
> safe to reclaim @addr. RCU tracks reader "transactions" begin/end,
> and makes sure that all pre-existing transactions are gone before
> synchronize_rcu() is allowed to complete. HP does this by tracking
> "posted" pointer slots with a HP domain. As long as hp_scan observes
> that HP readers are showing interest in @addr, it will wait.
> 
> One notable difference between RCU and HP is that HP knows exactly
> which pointer is blocking progress, and from which CPU (at least
> with my per-CPU HP domain implementation). Therefore, it is possible
> for HP to issue an IPI and make sure the HP user either completes its
> use of the pointer quickly, or stops using it right away (e.g. making
> the active mm use idle mm instead).
> 
> One strength of RCU is that it can track use of a whole set of RCU
> pointers just by tracking reader transaction begin/end, but this is
> also one of its weaknesses: a long reader transaction can postpone
> completion of grace period for a long time and increase the memory
> footprint. In comparison, HP can immediately complete as soon as the
> pointer it is scanning for is gone. Even better, it can send an IPI
> to the belate CPU and abort use of the pointer using a callback.

Plus, in contrast to hazard pointers, rcu_dereference() cannot say "no".

This all sounds like arguments *for* use of the term "publish" for
hazard pointers rather than against it.  What am I missing here?

							Thanx, Paul

> > > > > +/*
> > > > > + * hp_dereference_allocate: Dereference and allocate a hazard pointer.
> > > > > + *
> > > > > + * Returns a hazard pointer context. Expects to be called from preempt
> > > > > + * disable context.
> > > > > + */
> > > > 
> > > > More terrible naming. Same as above, but additionally, I would expect a
> > > > 'dereference' to actually dereference the pointer and have a return
> > > > value of the dereferenced type.
> > > 
> > > hp_dereference_try_post() ?
> > > 
> > > > 
> > > > This function seems to double check and update the hp_ctx thing. I'm not
> > > > at all sure yet wtf this is doing -- and the total lack of comments
> > > > aren't helping.
> > > 
> > > The hp_ctx contains the outputs.
> > > 
> > > The function loads *addr_p to then try_post it into a HP slot. On success,
> > > it re-reads the *addr_p (with address dependency) and if it still matches,
> > > use that as output address pointer.
> > > 
> > > I'm planning to remove hp_ctx, and just have:
> > > 
> > > /*
> > >   * hp_try_post: Try to post a hazard pointer.
> > >   *
> > >   * Post a hazard pointer slot for @addr. The object existence should
> > >   * be guaranteed by the caller. Expects to be called from preempt
> > >   * disable context.
> > >   *
> > >   * Returns true if post succeeds, false otherwise.
> > >   */
> > > static inline
> > > bool hp_try_post(struct hp_domain *hp_domain, void *addr, struct hp_slot **_slot)
> > > [...]
> > > 
> > > /*
> > >   * hp_dereference_try_post: Dereference and try to post a hazard pointer.
> > >   *
> > >   * Returns a hazard pointer context. Expects to be called from preempt
> > >   * disable context.
> > >   */
> > > static inline
> > > void *__hp_dereference_try_post(struct hp_domain *hp_domain,
> > >                                  void * const * addr_p, struct hp_slot **_slot)
> > > [...]
> > > 
> > > #define hp_dereference_try_post(domain, p, slot_p)              \
> > >          ((__typeof__(*(p))) __hp_dereference_try_post(domain, (void * const *) p, slot_p))
> > 
> > This will compile, but do the wrong thing when p is a regular pointer, no?
> 
> Right, at least in some cases the compiler may not complain, and people used to
> rcu_dereference() will expect that "p" is the pointer to load rather than the
> address of that pointer. This would be unexpected.
> 
> I must admit that passing the address holding the pointer to load rather than
> the pointer to load itself makes it much less troublesome in terms of macro
> layers. But perhaps this is another example where we should wander away from the
> beaten path and use a word different from "dereference" here. E.g.:
> 
> /*
>  * Use a comma expression within typeof: __typeof__((void)**(addr_p), *(addr_p))
>  * to generate a compile error if addr_p is not a pointer to a pointer.
>  */
> #define hp_load_try_post(domain, addr_p, slot_p)                \
>         ((__typeof__((void)**(addr_p), *(addr_p))) __hp_load_try_post(domain, (void * const *) (addr_p), slot_p))
> 
> > 
> > > 
> > > /* Clear the hazard pointer in @slot. */
> > > static inline
> > > void hp_remove(struct hp_slot *slot)
> > > [...]
> > 
> > Differently weird, but better I suppose :-)
> 
> If you find a better word than "remove" to pair with "post", I'm all in :)
> 
> > 
> > 
> > > > > +void hp_scan(struct hp_slot __percpu *percpu_slots, void *addr,
> > > > > +	     void (*retire_cb)(int cpu, struct hp_slot *slot, void *addr))
> > > > > +{
> > > > > +	int cpu;
> > > > > +
> > > > > +	/*
> > > > > +	 * Store A precedes hp_scan(): it unpublishes addr (sets it to
> > > > > +	 * NULL or to a different value), and thus hides it from hazard
> > > > > +	 * pointer readers.
> > > > > +	 */
> > > > > +
> > > > > +	if (!addr)
> > > > > +		return;
> > > > > +	/* Memory ordering: Store A before Load B. */
> > > > > +	smp_mb();
> > > > > +	/* Scan all CPUs slots. */
> > > > > +	for_each_possible_cpu(cpu) {
> > > > > +		struct hp_slot *slot = per_cpu_ptr(percpu_slots, cpu);
> > > > > +
> > > > > +		if (retire_cb && smp_load_acquire(&slot->addr) == addr)	/* Load B */
> > > > > +			retire_cb(cpu, slot, addr);
> > > > 
> > > > Is retirce_cb allowed to cmpxchg the thing?
> > > 
> > > It could, but we'd need to make sure the slot is not re-used by another
> > > hp_try_post() before the current user removes its own post. It would
> > > need to synchronize with the current HP user (e.g. with IPI).
> > > 
> > > I've actually renamed retire_cb to "on_match_cb".
> > 
> > Hmm, I think I see. Would it make sense to pass the expected addr to
> > hp_remove() and double check we don't NULL out something unexpected? --
> > maybe just for a DEBUG option.
> > 
> > I'm always seeing the NOHZ_FULL guys hating on this :-)
> 
> That's a fair point. Sure, we can do this as an extra safety net. For now I
> will just make the check always present, we can always move it to a debug
> option later.
> 
> And now I notice that hp_remove is also used for CPU hotplug (grep
> matches for cpuhp_remove_state()). I wonder if we should go for something
> more grep-friendly than "hp_", e.g. "hazptr_" and rename hp.h to hazptr.h ?
> 
> Thanks,
> 
> Mathieu
> 
> 
> -- 
> Mathieu Desnoyers
> EfficiOS Inc.
> https://www.efficios.com
>