Message-ID: <Pine.LNX.4.58.0802041153420.4119@gandalf.stny.rr.com>
Date:	Mon, 4 Feb 2008 12:09:00 -0500 (EST)
From:	Steven Rostedt <rostedt@...dmis.org>
To:	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
cc:	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	LKML <linux-kernel@...r.kernel.org>, Ingo Molnar <mingo@...e.hu>,
	Linus Torvalds <torvalds@...ux-foundation.org>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Christoph Hellwig <hch@...radead.org>,
	Mathieu Desnoyers <mathieu.desnoyers@...ymtl.ca>,
	Gregory Haskins <ghaskins@...ell.com>,
	Arnaldo Carvalho de Melo <acme@...stprotocols.net>,
	Thomas Gleixner <tglx@...utronix.de>,
	Tim Bird <tim.bird@...sony.com>,
	Sam Ravnborg <sam@...nborg.org>,
	"Frank Ch. Eigler" <fche@...hat.com>,
	Jan Kiszka <jan.kiszka@...mens.com>,
	John Stultz <johnstul@...ibm.com>,
	Arjan van de Ven <arjan@...radead.org>,
	Steven Rostedt <srostedt@...hat.com>
Subject: Re: [PATCH 02/22 -v7] Add basic support for gcc profiler instrumentation


Hi Paul,

First I want to say, "Thank you", for taking the time to explain this in
considerable detail. But I still have some minor questions.

 (Even though you already convinced me, but I still want full
  understanding ;-)


On Sat, 2 Feb 2008, Paul E. McKenney wrote:

> Yep, you have dependencies, so something like the following:
>
> initial state:
>
> 	struct foo {
> 		int a;
> 	};
> 	struct foo x = { 0 };
> 	struct foo y = { 0 };
> 	struct foo *global_p = &y;
> 	/* other variables are appropriately declared auto variables */
>
> 	/* No kmalloc() or kfree(), hence no RCU grace periods. */
> 	/* In the terminology of http://lwn.net/Articles/262464/, we */
> 	/* are doing only publish-subscribe, nothing else. */
>
> writer:
>
> 	x.a = 1;
> 	smp_wmb();  /* or smp_mb() */
> 	global_p = &x;
>
> reader:
>
> 	p = global_p;
> 	ta = p->a;
>
> Both Alpha and aggressive compiler optimizations can result in the reader
> seeing the new value of the pointer (&x) but the old value of the field
> (0).  Strange but true.  The fix is as follows:
>
> reader:
>
> 	p = global_p;
> 	smp_read_barrier_depends();  /* or use rcu_dereference() */
> 	ta = p->a;
>
> So how can this happen?  First note that if smp_read_barrier_depends()
> was unnecessary in this case, it would be unnecessary in all cases.
>
> Second, let's start with the compiler.  Suppose that a highly optimizing
> compiler notices that in almost all cases, the reader finds p==global_p.
> Suppose that this compiler also notices that one of the registers (say
> r1) almost always contains this expected value of global_p, and that
> cache pressure ensures that an actual load from global_p almost always
> generates an expensive cache miss.  Such a compiler would be within its
> rights (as defined by the C standard) to generate code assuming that r1
> already had the right value, while also generating code to validate this
> assumption, perhaps as follows:
>
> 	r2 = global_p;  /* high latency, other things complete meanwhile */
> 	ta == r1->a;
> 	if (r1 != r2)
> 		ta = r2->a;
>
> Now consider the following sequence of events on a superscalar CPU:

I think you missed one step here (causing my confusion). I don't want to
assume so I'll try to put in the missing step:

	writer: r1 = p;  /* happens to use r1 to store parameter p */

> 	reader: r2 = global_p; /* issued, has not yet completed. */
> 	reader: ta = r1->a; /* which gives zero. */
> 	writer: x.a = 1;
> 	writer: smp_wmb();
> 	writer: global_p = &x;
> 	reader: r2 = global_p; /* this instruction now completes */
> 	reader: if (r1 != r2) /* and these are equal, so we keep bad ta! */

Is that the case?

>
> I have great sympathy with the argument that this level of optimization
> is simply insane, but the fact is that there are real-world compilers
> that actually do this sort of thing.  In addition, there are cases where
> the compiler might be able to figure out that a value is constant, thus
> breaking the dependency chain.  This is most common for array references
> where the compiler might be able to figure out that a given array index
> is always zero, thus optimizing away the load and the dependency that
> the programmer might expect to enforce ordering.  (I have an example
> of this down at the end.)
>
> This sort of misordering is also done by DEC Alpha hardware, assuming
> split caches.  This can happen if the variable x is in an odd-numbered
> cache line and the variable global_p is in an even-numbered cache line.
> In this case, the smp_wmb() affects the memory order, but only within
> the writing CPU.  The ordering can be defeated in the reading CPU as
> follows:
>
> 	writer: x.a = 1;
> 	writer: smp_wmb();
> 	writer: global_p = &x;
> 	reader: p = global_p;
> 	reader: ta = p->a;
>
> 		But the reader's odd-numbered cache shard is loaded
> 		down with many queued cacheline invalidation requests,
> 		so the old cached version of x.a==0 remains in the
> 		reader's cache, so that the reader sees ta==0.
>
> In contrast:
>
> 	writer: x.a = 1;
> 	writer: smp_wmb();
> 	writer: global_p = &x;
> 	reader: p = global_p;
> 	reader: smp_read_barrier_depends();
>
> 		The above barrier forces all cacheline invalidation
> 		requests that have arrived at the reading CPU to be
> 		processed  before any subsequent reads, including
> 		the pending invalidation request for the variable x.
>
> 	reader: ta = p->a;
>
> 		So ta is now guaranteed to be 1, as desired.

Thanks, this is starting to clear things up for me (And scare me away from
Alpha's)

>
> > > > > Let me explain the situation here.
> > > > >
> > > > > We have a single link list called mcount_list that is walked when more
> > > > > than one function is registered by mcount. Mcount is called at the start
> > > > > of all C functions that are not annotated with "notrace". When more than
> > > > > one function is registered, mcount calls a loop function that does the
> > > > > following:
> > > > >
> > > > > notrace void mcount_list_func(unsigned long ip, unsigned long parent_ip)
> > > > > {
> > > > >         struct mcount_ops *op = mcount_list;
> > > >
> > > > When thinking RCU, this would be rcu_dereference and imply a read
> > > > barrier.
> > > >
> > > > >         while (op != &mcount_list_end) {
> > > > >                 op->func(ip, parent_ip);
> > > > >                 op = op->next;
> > > >
> > > > Same here; the rcu_dereference() would do the read depend barrier.
> > >
> > > Specifically:
> > >
> > > notrace void mcount_list_func(unsigned long ip, unsigned long parent_ip)
> > > {
> > >         struct mcount_ops *op = rcu_dereference(mcount_list);
> > >
> > >         while (op != &mcount_list_end) {
> > >                 op->func(ip, parent_ip);
> > >                 op = rcu_dereference(op->next);
> > >
> > > This assumes that you are using call_rcu(), synchronize_rcu(), or
> > > whatever to defer freeing/reuse of the ops structure.
> >
> > One special part of this is that the ops structure is never to be freed
> > (this is documented). It should be a static read-mostly structure.
> > Since it is not to be freed, I did not export the registered functions to
> > keep modules from using it. I may later add an export that will cause the
> > module to increment it's usage count so that it may never be freed.
> >
> > There's no guarantees that prevent the func from being called after it was
> > unregistered, nor should the users of this, ever touch the "next" pointer.
> >
> > This makes things easy when you don't need to free ;-)
>
> It can indeed make things easier, but it does not help in this case.
> This memory-ordering problem appears even if you never free anything, as
> described above.  Again, in the terminology laid out in the LWN article
> at http://lwn.net/Articles/262464/, you are doing a publish-subscribe
> operation, and it still must be protected.
>
> But yes, my comment above about using call_rcu() and friends did in fact
> incorrectly assume that you were freeing (or otherwise re-using) the
> data structures.
>
> > > > >         };
> > > > > }
> > > > >
> > > > > A registered function must already have a "func" filled, and the mcount
> > > > > register code takes care of "next".  It is documented that the calling
> > > > > function should "never" change next and always expect that the func can be
> > > > > called after it is unregistered. That's not the issue here.
> > > > >
> > > > > The issue is how to insert the ops into the list. I've done the following,
> > > > > as you can see in the code this text is inserted between.
> > > > >
> > > > >    ops->next = mcount_list;
> > > > >    smp_wmb();
> > > > >    mcount_list = ops;
> > > > >
> > > > > The read side pair is the reading of ops to ops->next, which should imply
> > > > > a smp_rmb() just by the logic. But Peter tells me things like alpha is
> > > > > crazy enough to do better than that! Thus, I'm asking you.
> > >
> > > Peter is correct when he says that Alpha does not necessarily respect data
> > > dependencies.  See the following URL for the official story:
> > >
> > > 	http://www.openvms.compaq.com/wizard/wiz_2637.html
> > >
> > > And I give an example hardware cache design that can result in this
> > > situation here:
> > >
> > > 	http://www.rdrop.com/users/paulmck/scalability/paper/ordering.2007.09.19a.pdf
> > >
> > > See the discussion starting with the "Why Reorder Memory Accesses?"
> > > heading in the second column of the first page.
> > >
> > > Strange, but true.  It took an Alpha architect quite some time to
> > > convince me of this back in the late 90s.  ;-)
> > >
> > > > > Can some arch have a reader where it receives ops->next before it received
> > > > > ops? This seems to me to be a phsyic arch, to know where ops->next is
> > > > > before it knows ops!
> > >
> > > The trick is that the machine might have a split cache, with (say)
> > > odd-numbered cache lines being processed by one half and even-numbered
> > > lines processed by the other half.  If reading CPU has one half of the
> > > cache extremely busy (e.g., processing invalidation requests from other
> > > CPUs) and the other half idle, memory misordering can happen in the
> > > receiving CPU -- if the pointer is processed by the idle half, and
> > > the pointed-to struct by the busy half, you might see the unitialized
> > > contents of the pointed-to structure.  The reading CPU must execute
> > > a memory barrier to force ordering in this case.
> > >
> > > > > Remember, that the ops that is being registered, is not viewable by any
> > > > > other CPU until mcount_list = ops. I don't see the need for a read barrier
> > > > > in this case. But I could very well be wrong.
> > >
> > > And I was right there with you before my extended discussions with the
> > > aforementioned Alpha architect!
> >
> > hmm, I'm still not convinced ;-)
> >
> > This is a unique situation. We don't need to worry about items being freed
> > because there's too many races to allow that. The items are only to
> > register functions and are not to be dynamically allocated or freed. In
> > this situation we do not need to worry about deletions.
> >
> > The smp_wmb is only for initialization of something that is about to enter
> > the list. It is not to protect against freeing.
>
> Similarly, the smp_read_barrier_depends() is only for initialization
> of something that is about to enter the list.  As with the smp_wmb()
> primitive, smp_read_barrier_depends() also is not to protect against
> freeing.  Instead, it is rcu_read_lock() and rcu_read_unlock() that
> protect against freeing.
>
> > Specifically:
> >
> >    ops->next = mcount_list;
> >    smp_wmb();
> >    mcount_list = ops;
> >
> > What this is to prevent is a new item that has next = NULL being viewable
> > to other CPUS before next is initalized.
>
> Were it not for aggressive compiler optimizations and DEC Alpha, you would
> be correct.  What this instead does is to do the writer's part of the job
> of preventing such new items from being visible to other CPUs before ->next
> is initialized.  These other CPUs must do their part as well, and that
> part is smp_read_barrier_depends() -- or rcu_dereference(), whichever is
> most appropriate.

Since the code doesn't use RCU, I'll keep with the
smp_read_barrier_depends().

>
> > On another cpu we have (simplified by removing loop):
> >
> >   op = mcount_list;
> >   op->func();
> >   op = op->next;
> >   if (op->next != NULL)
> >      op->func;
> >
> > What we want to prevent is reading of the new ops before ops->next is set.
>
> Understood.
>
> > What you are saying is that on alpha, even though the write to ops->next
> > has completed before mcount_list is set, we can still get a reversed
> > order?
>
> That is exactly what I am saying.  In addition, I am saying that
> aggressive compiler optimizations can have this same effect, even on
> non-Alpha CPUs.
>
> >   ops->next = mcount_list;  -- in one cache line
> >   smp_wmb();
> >   mcount_list = ops;       -- in another cache line
> >
> > Even though the ops->next is completed, we can have on another cpu:
> >
> >    op = mcount_list; (which is the ops from above)
> >    op = op->next;  -- still see the old ops->next?
>
> Yes, this bizarre sequence of events really can happen.  The fix is to
> do the following:
>
>    op = mcount_list; (which is the ops from above)
>    smp_read_barrier_depends();
>    op = op->next;  -- no longer see the old ops->next
>
> > I just want to understand this. I already put in the read_barrier_depends
> > because it doesn't hurt on most archs anyway (nops).
>
> Very good!!!
>
> And here is the example using array indexes.
>
> initial state:
>
> 	struct foo {
> 		int a;
> 	};
> 	struct foo x[ARRAY_SIZE] = { 0 };
> 	struct foo *global_p = &x[0];
> 	/* other variables are appropriately declared auto variables */
>
> 	/* No kmalloc() or kfree(), hence no RCU grace periods. */
> 	/* In the terminology of http://lwn.net/Articles/262464/, we */
> 	/* are doing only publish-subscribe, nothing else. */
>
> writer:
>
> 	x[cur_idx].a = 1;
> 	smp_wmb();  /* or smp_mb() */
> 	global_idx = cur_idx;
>
> reader:
>
> 	i = global_idx;
> 	ta = x[i].a
>
> Suppose we have ARRAY_SIZE of 1.  Then the standard states that the
> results of indexing x[] with a non-zero index are undefined.  Since they
> are undefined, the compiler is within its rights to assume that the
> index will always be zero, so that the reader code would be as follows:
>
> reader:
>
> 	ta = x[0].a
>
> No dependency, no ordering.  So this totally reasonable generated code
> could see the pre-initialized value of field a.  The job of both
> smp_read_barrier_depends() and rcu_dereference() is to tell both the
> CPU and the compiler that such assumptions are ill-advised.
>

Paul,

Thanks again for this lengthy email. It took me several readings to absorb
it all in.

I recommend that someone have a pointer to this email because it really
does explain why read_barrier_depends is needed.

Excellent job of explaining this!!! Much appreciated.

-- Steve

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists