Make use of arch_mutex_cpu_relax() so architectures can override the
default cpu_relax() semantics.
This is especially useful for s390, where cpu_relax() means that we
yield() the current (virtual) cpu and therefore is very expensive,
and would contradict the whole purpose of the lockless cmpxchg loop.
Signed-off-by: NHeiko Carstens <heiko.carstens@de.ibm.com>

The 64-bit cmpxchg operation on the lockref is ordered by virtue of
hazarding between the cmpxchg operation and the reference count
manipulation. On weakly ordered memory architectures (such as ARM), it
can be of great benefit to omit the barrier instructions where they are
not needed.

This patch moves the lockless lockref code over to a cmpxchg64_relaxed
operation, which doesn't provide barrier semantics. If the operation
isn't defined, we simply #define it as the usual 64-bit cmpxchg macro.

Cc: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: NWill Deacon <will.deacon@arm.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

The cmpxchg() function tends not to support 64-bit arguments on 32-bit
architectures.  This could be either due to use of unsigned long
arguments (like on ARM) or lack of instruction support (cmpxchgq on
x86).  However, these architectures may implement a specific cmpxchg64()
function to provide 64-bit cmpxchg support instead.

Since the lockref code requires a 64-bit cmpxchg and relies on the
architecture selecting ARCH_USE_CMPXCHG_LOCKREF, move to using cmpxchg64
instead of cmpxchg and allow 32-bit architectures to make use of the
lockless lockref implementation.

Cc: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: NWill Deacon <will.deacon@arm.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

The only actual current lockref user (dcache) uses zero reference counts
even for perfectly live dentries, because it's a cache: there may not be
any users, but that doesn't mean that we want to throw away the dentry.

At the same time, the dentry cache does have a notion of a truly "dead"
dentry that we must not even increment the reference count of, because
we have pruned it and it is not valid.

Currently that distinction is not visible in the lockref itself, and the
dentry cache validation uses "lockref_get_or_lock()" to either get a new
reference to a dentry that already had existing references (and thus
cannot be dead), or get the dentry lock so that we can then verify the
dentry and increment the reference count under the lock if that
verification was successful.

That's all somewhat complicated.

This adds the concept of being "dead" to the lockref itself, by simply
using a count that is negative.  This allows a usage scenario where we
can increment the refcount of a dentry without having to validate it,
and pushing the special "we killed it" case into the lockref code.

The dentry code itself doesn't actually use this yet, and it's probably
too late in the merge window to do that code (the dentry_kill() code
with its "should I decrement the count" logic really is pretty complex
code), but let's introduce the concept at the lockref level now.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

The code got rewritten, but the comments got copied as-is from older
versions, and as a result the argument name in the comment didn't
actually match the code any more.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

While we are likley to succeed and break out of this loop, it isn't
guaranteed.  We should be power and thread friendly if we do have to
go around for a second (or third, or more) attempt.
Signed-off-by: NTony Luck <tony.luck@intel.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Instead of taking the spinlock, the lockless versions atomically check
that the lock is not taken, and do the reference count update using a
cmpxchg() loop.  This is semantically identical to doing the reference
count update protected by the lock, but avoids the "wait for lock"
contention that you get when accesses to the reference count are
contended.

Note that a "lockref" is absolutely _not_ equivalent to an atomic_t.
Even when the lockref reference counts are updated atomically with
cmpxchg, the fact that they also verify the state of the spinlock means
that the lockless updates can never happen while somebody else holds the
spinlock.

So while "lockref_put_or_lock()" looks a lot like just another name for
"atomic_dec_and_lock()", and both optimize to lockless updates, they are
fundamentally different: the decrement done by atomic_dec_and_lock() is
truly independent of any lock (as long as it doesn't decrement to zero),
so a locked region can still see the count change.

The lockref structure, in contrast, really is a *locked* reference
count.  If you hold the spinlock, the reference count will be stable and
you can modify the reference count without using atomics, because even
the lockless updates will see and respect the state of the lock.

In order to enable the cmpxchg lockless code, the architecture needs to
do three things:

 (1) Make sure that the "arch_spinlock_t" and an "unsigned int" can fit
     in an aligned u64, and have a "cmpxchg()" implementation that works
     on such a u64 data type.

 (2) define a helper function to test for a spinlock being unlocked
     ("arch_spin_value_unlocked()")

 (3) select the "ARCH_USE_CMPXCHG_LOCKREF" config variable in its
     Kconfig file.

This enables it for x86-64 (but not 32-bit, we'd need to make sure
cmpxchg() turns into the proper cmpxchg8b in order to enable it for
32-bit mode).
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

They aren't very good to inline, since they already call external
functions (the spinlock code), and we're going to create rather more
complicated versions of them that can do the reference count updates
locklessly.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

This behaves like "lockref_get_not_zero()", but instead of doing nothing
if the count was zero, it returns with the lock held.

This allows callers to revalidate the lockref-protected data structure
if required even if the count was zero to begin with, and possibly
increment the count if it passes muster.

In particular, the dentry code wants this when it wants to turn an
RCU-protected dentry into a stable refcounted one: if the dentry count
it zero, but the sequence number still validates the dentry, we can take
a reference to it.
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

This introduces a new "lockref" structure that supports the concept of
lockless updates of reference counts that still honor an attached
spinlock.

NOTE! This reference implementation is not the optimized lockless
version, rather it is the fallback implementation using standard
spinlocks.  The actual optimized versions will be merged into 3.12, but
I wanted to get the infrastructure in place and document the new
interfaces.

[ Also note that this particular commit is drastically cut-down minimal
  version of the original patch by Waiman.  In order to properly credit
  the original author I'm marking Waiman as the author here, but in the
  end this patch bears little resemblance to the patch by Waiman.  So
  blame any errors on me editing things down to the point where I can
  introduce the infrastructure before the merge window for 3.12 actually
  opens.     - Linus ]
Signed-off-by: NWaiman Long <Waiman.Long@hp.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>