linux_dsm_epyc7002/Documentation/arm/vlocks.txt

vlocks for Bare-Metal Mutual Exclusion
======================================

Voting Locks, or "vlocks" provide a simple low-level mutual exclusion
mechanism, with reasonable but minimal requirements on the memory
system.

These are intended to be used to coordinate critical activity among CPUs
which are otherwise non-coherent, in situations where the hardware
provides no other mechanism to support this and ordinary spinlocks
cannot be used.


vlocks make use of the atomicity provided by the memory system for
writes to a single memory location.  To arbitrate, every CPU "votes for
itself", by storing a unique number to a common memory location.  The
final value seen in that memory location when all the votes have been
cast identifies the winner.

In order to make sure that the election produces an unambiguous result
in finite time, a CPU will only enter the election in the first place if
no winner has been chosen and the election does not appear to have
started yet.


Algorithm
---------

The easiest way to explain the vlocks algorithm is with some pseudo-code:


	int currently_voting[NR_CPUS] = { 0, };
	int last_vote = -1; /* no votes yet */

	bool vlock_trylock(int this_cpu)
	{
		/* signal our desire to vote */
		currently_voting[this_cpu] = 1;
		if (last_vote != -1) {
			/* someone already volunteered himself */
			currently_voting[this_cpu] = 0;
			return false; /* not ourself */
		}

		/* let's suggest ourself */
		last_vote = this_cpu;
		currently_voting[this_cpu] = 0;

		/* then wait until everyone else is done voting */
		for_each_cpu(i) {
			while (currently_voting[i] != 0)
				/* wait */;
		}

		/* result */
		if (last_vote == this_cpu)
			return true; /* we won */
		return false;
	}

	bool vlock_unlock(void)
	{
		last_vote = -1;
	}


The currently_voting[] array provides a way for the CPUs to determine
whether an election is in progress, and plays a role analogous to the
"entering" array in Lamport's bakery algorithm [1].

However, once the election has started, the underlying memory system
atomicity is used to pick the winner.  This avoids the need for a static
priority rule to act as a tie-breaker, or any counters which could
overflow.

As long as the last_vote variable is globally visible to all CPUs, it
will contain only one value that won't change once every CPU has cleared
its currently_voting flag.


Features and limitations
------------------------

 * vlocks are not intended to be fair.  In the contended case, it is the
   _last_ CPU which attempts to get the lock which will be most likely
   to win.

   vlocks are therefore best suited to situations where it is necessary
   to pick a unique winner, but it does not matter which CPU actually
   wins.

 * Like other similar mechanisms, vlocks will not scale well to a large
   number of CPUs.

   vlocks can be cascaded in a voting hierarchy to permit better scaling
   if necessary, as in the following hypothetical example for 4096 CPUs:

	/* first level: local election */
	my_town = towns[(this_cpu >> 4) & 0xf];
	I_won = vlock_trylock(my_town, this_cpu & 0xf);
	if (I_won) {
		/* we won the town election, let's go for the state */
		my_state = states[(this_cpu >> 8) & 0xf];
		I_won = vlock_lock(my_state, this_cpu & 0xf));
		if (I_won) {
			/* and so on */
			I_won = vlock_lock(the_whole_country, this_cpu & 0xf];
			if (I_won) {
				/* ... */
			}
			vlock_unlock(the_whole_country);
		}
		vlock_unlock(my_state);
	}
	vlock_unlock(my_town);


ARM implementation
------------------

The current ARM implementation [2] contains some optimisations beyond
the basic algorithm:

 * By packing the members of the currently_voting array close together,
   we can read the whole array in one transaction (providing the number
   of CPUs potentially contending the lock is small enough).  This
   reduces the number of round-trips required to external memory.

   In the ARM implementation, this means that we can use a single load
   and comparison:

	LDR	Rt, [Rn]
	CMP	Rt, #0

   ...in place of code equivalent to:

	LDRB	Rt, [Rn]
	CMP	Rt, #0
	LDRBEQ	Rt, [Rn, #1]
	CMPEQ	Rt, #0
	LDRBEQ	Rt, [Rn, #2]
	CMPEQ	Rt, #0
	LDRBEQ	Rt, [Rn, #3]
	CMPEQ	Rt, #0

   This cuts down on the fast-path latency, as well as potentially
   reducing bus contention in contended cases.

   The optimisation relies on the fact that the ARM memory system
   guarantees coherency between overlapping memory accesses of
   different sizes, similarly to many other architectures.  Note that
   we do not care which element of currently_voting appears in which
   bits of Rt, so there is no need to worry about endianness in this
   optimisation.

   If there are too many CPUs to read the currently_voting array in
   one transaction then multiple transations are still required.  The
   implementation uses a simple loop of word-sized loads for this
   case.  The number of transactions is still fewer than would be
   required if bytes were loaded individually.


   In principle, we could aggregate further by using LDRD or LDM, but
   to keep the code simple this was not attempted in the initial
   implementation.


 * vlocks are currently only used to coordinate between CPUs which are
   unable to enable their caches yet.  This means that the
   implementation removes many of the barriers which would be required
   when executing the algorithm in cached memory.

   packing of the currently_voting array does not work with cached
   memory unless all CPUs contending the lock are cache-coherent, due
   to cache writebacks from one CPU clobbering values written by other
   CPUs.  (Though if all the CPUs are cache-coherent, you should be
   probably be using proper spinlocks instead anyway).


 * The "no votes yet" value used for the last_vote variable is 0 (not
   -1 as in the pseudocode).  This allows statically-allocated vlocks
   to be implicitly initialised to an unlocked state simply by putting
   them in .bss.

   An offset is added to each CPU's ID for the purpose of setting this
   variable, so that no CPU uses the value 0 for its ID.


Colophon
--------

Originally created and documented by Dave Martin for Linaro Limited, for
use in ARM-based big.LITTLE platforms, with review and input gratefully
received from Nicolas Pitre and Achin Gupta.  Thanks to Nicolas for
grabbing most of this text out of the relevant mail thread and writing
up the pseudocode.

Copyright (C) 2012-2013  Linaro Limited
Distributed under the terms of Version 2 of the GNU General Public
License, as defined in linux/COPYING.


References
----------

[1] Lamport, L. "A New Solution of Dijkstra's Concurrent Programming
    Problem", Communications of the ACM 17, 8 (August 1974), 453-455.

    https://en.wikipedia.org/wiki/Lamport%27s_bakery_algorithm

[2] linux/arch/arm/common/vlock.S, www.kernel.org.