ACCESS_ONCE does not work reliably on non-scalar types. For
example gcc 4.6 and 4.7 might remove the volatile tag for such
accesses during the SRA (scalar replacement of aggregates) step
(https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58145)

Change the spinlock code to replace ACCESS_ONCE with READ_ONCE.

Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

When unlocking a spinlock, we require the following, strictly ordered
sequence of events:

	<barrier>	/* dmb */
	<unlock>
	<barrier>	/* dsb */
	<sev>

Whilst the code does indeed reflect this in terms of the architecture,
the final <barrier> + <sev> have been contracted into a single inline
asm without a "memory" clobber, therefore the compiler is at liberty to
reorder the unlock to the end of the above sequence. In such a case,
a waiting CPU may be woken up before the lock has been unlocked, leading
to extremely poor performance.

This patch reworks the dsb_sev() function to make use of the dsb()
macro and ensure ordering against the unlock.

Cc: <stable@vger.kernel.org>
Reported-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

Conflicts:
	arch/arm/include/asm/atomic.h
	arch/arm/include/asm/hardirq.h
	arch/arm/kernel/smp.c

Our spinlocks are only 32-bit (2x16-bit tickets) and, on processors
with 64-bit atomic instructions, cmpxchg64 makes use of the double-word
exclusive accessors.

This patch wires up the cmpxchg-based lockless lockref implementation
for ARM.

Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

The cost of changing a cacheline from shared to exclusive state can be
significant, especially when this is triggered by an exclusive store,
since it may result in having to retry the transaction.

This patch prefixes our {spin,read,write}_[try]lock implementations with
pldw instructions (on CPUs which support them) to try and grab the line
in exclusive state from the start. arch_rwlock_t is changed to avoid
using a volatile member, since this generates compiler warnings when
falling back on the __builtin_prefetch intrinsic which expects a const
void * argument.

Acked-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>

Patching UP/SMP alternatives inside inline assembly blocks is useful
outside of the spinlock implementation, where it is used for sev and wfe.

This patch lifts the macro into processor.h and gives it a scarier name
to (a) avoid conflicts in the global namespace and (b) to try and deter
its usage unless you "know what you're doing". The W macro for generating
wide instructions when targetting Thumb-2 is also made available under
the name WASM, to reduce the potential for conflicts with other headers.

Acked-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>

Commit 15e7e5c1ebf5 ("ARM: 7749/1: spinlock: retry trylock operation if
strex fails on free lock") modifying our arch_spin_trylock to retry the
acquisition if the lock appeared uncontended, but the strex failed.

This patch does the same for rwlocks, which were missed by the original
patch.

Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

The res variable is written before we've finished with the input
operands (namely the lock address), so ensure that we mark it as `early
clobber' to avoid unintended register sharing.

Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

When unlocking a spinlock, we use the sev instruction to signal other
CPUs waiting on the lock. Since sev is not a memory access instruction,
we require a dsb in order to ensure that the sev is not issued ahead
of the store placing the lock in an unlocked state.

However, as sev is only concerned with other processors in a
multiprocessor system, we can restrict the scope of the preceding dsb
to the inner-shareable domain. Furthermore, we can restrict the scope to
consider only stores, since there are no independent loads on the unlock
path.

A side-effect of this change is that a spin_unlock operation no longer
forces completion of pending TLB invalidation, something which we rely
on when unlocking runqueues to ensure that CPU migration during TLB
maintenance routines doesn't cause us to continue before the operation
has completed.

This patch adds the -ishst suffix to the ARMv7 definition of dsb_sev()
and adds an inner-shareable dsb to the context-switch path when running
a preemptible, SMP, v7 kernel.

Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>