Re: [PATCH 1/24] make atomic_read() behave consistently on alpha

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Paul E. McKenney wrote:
On Thu, Aug 09, 2007 at 03:24:40PM -0400, Chris Snook wrote:
Paul E. McKenney wrote:
On Thu, Aug 09, 2007 at 02:13:52PM -0400, Chris Snook wrote:
Paul E. McKenney wrote:
On Thu, Aug 09, 2007 at 01:14:35PM -0400, Chris Snook wrote:
If you're depending on volatile writes being visible to other CPUs, you're screwed either way, because the CPU can hold that data in cache as long as it wants before it writes it to memory. When this finally does happen, it will happen atomically, which is all that atomic_set guarantees. If you need to guarantee that the value is written to memory at a particular time in your execution sequence, you either have to read it from memory to force the compiler to store it first (and a volatile cast in atomic_read will suffice for this) or you have to use LOCK_PREFIX instructions which will invalidate remote cache lines containing the same variable. This patch doesn't change either of these cases.
The case that it -can- change is interactions with interrupt handlers.
And NMI/SMI handlers, for that matter.
You have a point here, but only if you can guarantee that the interrupt handler is running on a processor sharing the cache that has the not-yet-written volatile value. That implies a strictly non-SMP architecture. At the moment, none of those have volatile in their declaration of atomic_t, so this patch can't break any of them.
This can also happen when using per-CPU variables.  And there are a
number of per-CPU variables that are either atomic themselves or are
structures containing atomic fields.
Accessing per-CPU variables in this fashion reliably already requires a suitable smp/non-smp read/write memory barrier. I maintain that if we break anything with this change, it was really already broken, if less obviously. Can you give a real or synthetic example of legitimate code that could break?

My main concern is actually the lack of symmetry -- I would expect
that an atomic_set() would have the same properties as atomic_read().
It is easy and cheap to provide them with similar properties, so why not?
Debugging even a single problem would consume far more time than simply
giving them corresponding semantics.

But you asked for examples.  These are synthetic, and of course legitimacy
is in the eye of the beholder.

1.  Watchdog variable.

	atomic_t watchdog = ATOMIC_INIT(0);

	...

	int i;
	while (!done) {

		/* Do so stuff that doesn't take more than a few us. */
		/* Could do atomic increment, but throughput penalty. */

		i++;
		atomic_set(&watchdog, i);
	}
	do_something_with(&watchdog);


	/* Every so often on some other CPU... */

	if ((new_watchdog = atomic_read(&watchdog)) == old_watchdog)
		die_horribly();
	old_watchdog = new_watchdog;


	If atomic_set() did not have volatile semantics, the compiler
	would be within its rights optimizing it to simply get the
	final value of "i" after exit from the loop.  This would cause
	the watchdog check to fail spuriously.  Memory barriers are
	not required in this case, because the CPU cannot hang onto
	the value for very long -- we don't care about the exact value,
	or about exact synchronization, but rather about whether or
	not the value is changing.

	In this (toy) example, one might replace the atomic_set() with
	an atomic increment (though that might be too expensive in some
	cases) or with something like:

		atomic_set(&watchdog, atomic_read(&watchdog) + 1);

	However, other cases might not permit this transformation,
	for example, an existing heavily used API might take int rather
	than atomic_t.

	Some will no doubt argue that this example should use a
	macro or an asm similar to the "forget()" asm put forward
	elsewhere in this thread.

2.  Communicating both with interrupt handler and with other CPUs.
    For example, data elements that are built up in a location visible
    to interrupts and NMIs, and then added as a unit to a data structure
    visible to other CPUs.  This more-realistic example is abbreviated
    to the point of pointlessness as follows:

	struct foo {
		atomic_t a;
		atomic_t b;
	};

	DEFINE_PER_CPU(struct foo *, staging) = NULL;

	/* Create element in staging area. */

	__get_cpu_var(staging) = kzalloc(sizeof(*p), GFP_WHATEVER);
	if (__get_cpu_var(staging) == NULL)
		die_horribly();
	/* allocate an element of some per-CPU array, get the result in "i" */
	atomic_set(__get_cpu_var(staging).a, i);
	/* allocate another element of a per-CPU array, with result in "i" */
	atomic_set(__get_cpu_var(staging).b, i);
	rcu_assign_pointer(some_global_place, __get_cpu_var(staging));

	If atomic_set() didn't have volatile semantics, then an interrupt
	or NMI handler could see the atomic_set() to .a and .b out of
	order due to compiler optimizations.

Remember, you -did- ask for these!!!  ;-)

Ok, I'm convinced. Part of the motivation here is to avoid heisenbugs, so if people expect volatile atomic_set behavior, I'm inclined to give it to them. I don't really feel like indulging the compiler bug paranoiacs, but developer expectations are a legitimate motivation, and a major part of why I posted this in the first place. I'll resubmit the patchset with a volatile cast in atomic_set. Before I do, is there anything *else* that desperately needs such a cast? As far as I can tell, all the other functions are implemented with __asm__ __volatile__, or with spinlocks that use that under the hood.

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

[Index of Archives]     [Kernel Newbies]     [Netfilter]     [Bugtraq]     [Photo]     [Stuff]     [Gimp]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Video 4 Linux]     [Linux for the blind]     [Linux Resources]
  Powered by Linux