Annotate the release barrier and memory clobber (in effect, producing a
compiler barrier) in the publish_tail_cpu call. These barriers have the
effect of ensuring that qnode attributes are all written to prior to
publish the node to the waitqueue.
Even while the initial write to the 'locked' attribute is guaranteed to
terminate prior to the node being visible, KCSAN still complains that
the write is reorderable by the compiler. Issue a kcsan_release() to
inform KCSAN of the release barrier contained in publish_tail_cpu().
Signed-off-by: Rohan McLure <rmclure@linux.ibm.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20230510033117.1395895-3-rmclure@linux.ibm.com
The powerpc implementation of qspinlocks will both poll and spin on the
bitlock guarding a qnode. Mark these accesses with READ_ONCE to convey
to KCSAN that polling is intentional here.
Signed-off-by: Rohan McLure <rmclure@linux.ibm.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20230510033117.1395895-2-rmclure@linux.ibm.com
This adds compile-time options that allow the EH lock hint bit to be
enabled or disabled, and adds some new options that may or may not
help matters. To help with experimentation and tuning.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-18-npiggin@gmail.com
Finding the owner or a queued waiter on a lock with a preempted vcpu is
indicative of an oversubscribed guest causing the lock to get into
trouble. Provide some options to detect this situation and have new CPUs
avoid queueing for a longer time (more steal iterations) to minimise the
problems caused by vcpu preemption on the queue.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-17-npiggin@gmail.com
Provide an option that holds off queueing indefinitely while the lock
owner is preempted. This could reduce queueing latencies for very
overcommitted vcpu situations.
This is disabled by default.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-16-npiggin@gmail.com
Allow for a reduction in the number of times a CPU from a different
node than the owner can attempt to steal the lock before queueing.
This could bias the transfer behaviour of the lock across the
machine and reduce NUMA crossings.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-15-npiggin@gmail.com
Use the spin_begin/spin_cpu_relax/spin_end APIs in qspinlock, which helps
to prevent threads issuing a lot of expensive priority nops which may not
have much effect due to immediately executing low then medium priority.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-14-npiggin@gmail.com
This change allows trylock to steal the lock. It also allows the initial
lock attempt to steal the lock rather than bailing out and going to the
slow path.
This gives trylock more strength: without this a continually-contended
lock will never permit a trylock to succeed. With this change, the
trylock has a small but non-zero chance.
It also gives the lock fastpath most of the benefit of passing the
reservation back through to the steal loop in the slow path without the
complexity.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-13-npiggin@gmail.com
After the head of the queue acquires the lock, it releases the
next waiter in the queue to become the new head. Add an option
to prod the new head if its vCPU was preempted. This may only
have an effect if queue waiters are yielding.
Disable this option by default for now, i.e., no logical change.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-12-npiggin@gmail.com
Having all CPUs poll the lock word for the owner CPU that should be
yielded to defeats most of the purpose of using MCS queueing for
scalability. Yet it may be desirable for queued waiters to yield to a
preempted owner.
With this change, queue waiters never sample the owner CPU directly from
the lock word. The queue head (which is spinning on the lock) propagates
the owner CPU back to the next waiter if it finds the owner has been
preempted. That waiter then propagates the owner CPU back to the next
waiter, and so on.
s390 addresses this problem differenty, by having queued waiters sample
the lock word to find the owner at a low frequency. That has the
advantage of being simpler, the advantage of propagation is that the
lock word never has to be accesed by queued waiters, and the transfer of
cache lines to transmit the owner data is only required when lock holder
vCPU preemption occurs.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-11-npiggin@gmail.com
If the head of queue is preventing stealing but it finds the owner vCPU
is preempted, it will yield its cycles to the owner which could cause it
to become preempted. Add an option to re-allow stealers before yielding,
and disallow them again after returning from the yield.
Disable this option by default for now, i.e., no logical change.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-10-npiggin@gmail.com
Queued waiters which are not at the head of the queue don't spin on
the lock word but their qnode lock word, waiting for the previous queued
CPU to release them. Add an option which allows these waiters to yield
to the previous CPU if its vCPU is preempted.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-9-npiggin@gmail.com
Waiters spinning on the lock word should yield to the lock owner if the
vCPU is preempted. This improves performance when the hypervisor has
oversubscribed physical CPUs.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-8-npiggin@gmail.com
Store the owner CPU number in the lock word so it may be yielded to,
as powerpc's paravirtualised simple spinlocks do.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-7-npiggin@gmail.com
Give the queue head the ability to stop stealers. After a number of
spins without successfully acquiring the lock, the queue head sets
this, which halts stealing and will assure it is the next owner.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-6-npiggin@gmail.com
Allow new waiters to "steal" the lock before queueing. That is, to
acquire it while other CPUs have queued.
This particularly helps paravirt performance when physical CPUs are
oversubscribed, by keeping the lock from becoming a strict FIFO and
vCPU preemption causing queue train wrecks.
The new __queued_spin_trylock_steal() function is put in qspinlock.h
to save having to move it, because it will be used there by a later
change.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-5-npiggin@gmail.com
This uses more optimal ll/sc style access patterns (rather than
cmpxchg), and also sets the EH=1 lock hint on those operations
which acquire ownership of the lock.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-4-npiggin@gmail.com
This forms the basis of the qspinlock slow path.
Like generic qspinlocks and unlike the vanilla MCS algorithm, the lock
owner does not participate in the queue, only waiters. The first waiter
spins on the lock word, then when the lock is released it takes
ownership and unqueues the next waiter. This is how qspinlocks can be
implemented with the spinlock API -- lock owners don't need a node, only
waiters do.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-2-npiggin@gmail.com
Add a powerpc specific implementation of queued spinlocks. This is the
build framework with a very simple (non-queued) spinlock implementation
to begin with. Later changes add queueing, and other features and
optimisations one-at-a-time. It is done this way to more easily see how
the queued spinlocks are built, and to make performance and correctness
bisects more useful.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Drop paravirt.h & processor.h changes to fix 32-bit build]
[mpe: Fix 32-bit build of qspinlock.o & disallow GENERIC_LOCKBREAK per Nick]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/CONLLQB6DCJU.2ZPOS7T6S5GRR@bobo