We'd like all architectures to convert to ARCH_ATOMIC, as once all
architectures are converted it will be possible to make significant
cleanups to the atomics headers, and this will make it much easier to
generically enable atomic functionality (e.g. debug logic in the
instrumented wrappers).
As a step towards that, this patch migrates powerpc to ARCH_ATOMIC. The
arch code provides arch_{atomic,atomic64,xchg,cmpxchg}*(), and common
code wraps these with optional instrumentation to provide the regular
functions.
While atomic_try_cmpxchg_lock() is not part of the common atomic API, it
is given an `arch_` prefix for consistency.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210525140232.53872-28-mark.rutland@arm.com
49a7d46a06 (powerpc: Implement smp_cond_load_relaxed()) added
busy-waiting pausing with a preferred SMT priority pattern, lowering
the priority (reducing decode cycles) during the whole loop slowpath.
However, data shows that while this pattern works well with simple
spinlocks, queued spinlocks benefit more being kept in medium priority,
with a cpu_relax() instead, being a low+medium combo on powerpc.
Data is from three benchmarks on a Power9: 9008-22L 64 CPUs with
2 sockets and 8 threads per core.
1. locktorture.
This is data for the lowest and most artificial/pathological level,
with increasing thread counts pounding on the lock. Metrics are total
ops/minute. Despite some small hits in the 4-8 range, scenarios are
either neutral or favorable to this patch.
+=========+==========+==========+=======+
| # tasks | vanilla | dirty | %diff |
+=========+==========+==========+=======+
| 2 | 46718565 | 48751350 | 4.35 |
+---------+----------+----------+-------+
| 4 | 51740198 | 50369082 | -2.65 |
+---------+----------+----------+-------+
| 8 | 63756510 | 62568821 | -1.86 |
+---------+----------+----------+-------+
| 16 | 67824531 | 70966546 | 4.63 |
+---------+----------+----------+-------+
| 32 | 53843519 | 61155508 | 13.58 |
+---------+----------+----------+-------+
| 64 | 53005778 | 53104412 | 0.18 |
+---------+----------+----------+-------+
| 128 | 53331980 | 54606910 | 2.39 |
+=========+==========+==========+=======+
2. sockperf (tcp throughput)
Here a client will do one-way throughput tests to a localhost server, with
increasing message sizes, dealing with the sk_lock. This patch shows to put
the performance of the qspinlock back to par with that of the simple lock:
simple-spinlock vanilla dirty
Hmean 14 73.50 ( 0.00%) 54.44 * -25.93%* 73.45 * -0.07%*
Hmean 100 654.47 ( 0.00%) 385.61 * -41.08%* 771.43 * 17.87%*
Hmean 300 2719.39 ( 0.00%) 2181.67 * -19.77%* 2666.50 * -1.94%*
Hmean 500 4400.59 ( 0.00%) 3390.77 * -22.95%* 4322.14 * -1.78%*
Hmean 850 6726.21 ( 0.00%) 5264.03 * -21.74%* 6863.12 * 2.04%*
3. dbench (tmpfs)
Configured to run with up to ncpusx8 clients, it shows both latency and
throughput metrics. For the latency, with the exception of the 64 case,
there is really nothing to go by:
vanilla dirty
Amean latency-1 1.67 ( 0.00%) 1.67 * 0.09%*
Amean latency-2 2.15 ( 0.00%) 2.08 * 3.36%*
Amean latency-4 2.50 ( 0.00%) 2.56 * -2.27%*
Amean latency-8 2.49 ( 0.00%) 2.48 * 0.31%*
Amean latency-16 2.69 ( 0.00%) 2.72 * -1.37%*
Amean latency-32 2.96 ( 0.00%) 3.04 * -2.60%*
Amean latency-64 7.78 ( 0.00%) 8.17 * -5.07%*
Amean latency-512 186.91 ( 0.00%) 186.41 * 0.27%*
For the dbench4 Throughput (misleading but traditional) there's a small
but rather constant improvement:
vanilla dirty
Hmean 1 849.13 ( 0.00%) 851.51 * 0.28%*
Hmean 2 1664.03 ( 0.00%) 1663.94 * -0.01%*
Hmean 4 3073.70 ( 0.00%) 3104.29 * 1.00%*
Hmean 8 5624.02 ( 0.00%) 5694.16 * 1.25%*
Hmean 16 9169.49 ( 0.00%) 9324.43 * 1.69%*
Hmean 32 11969.37 ( 0.00%) 12127.09 * 1.32%*
Hmean 64 15021.12 ( 0.00%) 15243.14 * 1.48%*
Hmean 512 14891.27 ( 0.00%) 15162.11 * 1.82%*
Measuring the dbench4 Per-VFS Operation latency, shows some very minor
differences within the noise level, around the 0-1% ranges.
Fixes: 49a7d46a06 ("powerpc: Implement smp_cond_load_relaxed()")
Acked-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210318204702.71417-1-dave@stgolabs.net
c6f5d02b6a (locking/spinlocks/arm64: Remove smp_mb() from
arch_spin_is_locked()) made it pretty official that the call
semantics do not imply any sort of barriers, and any user that
gets creative must explicitly do any serialization.
This creativity, however, is nowadays pretty limited:
1. spin_unlock_wait() has been removed from the kernel in favor
of a lock/unlock combo. Furthermore, queued spinlocks have now
for a number of years no longer relied on _Q_LOCKED_VAL for the
call, but any non-zero value to indicate a locked state. There
were cases where the delayed locked store could lead to breaking
mutual exclusion with crossed locking; such as with sysv ipc and
netfilter being the most extreme.
2. The auditing Andrea did in verified that remaining spin_is_locked()
no longer rely on such semantics. Most callers just use it to assert
a lock is taken, in a debug nature. The only user that gets cute is
NOLOCK qdisc, as of:
96009c7d50 (sched: replace __QDISC_STATE_RUNNING bit with a spin lock)
... which ironically went in the next day after c6f5d02b6a. This
change replaces test_bit() with spin_is_locked() to know whether
to take the busylock heuristic to reduce contention on the main
qdisc lock. So any races against spin_is_locked() for archs that
use LL/SC for spin_lock() will be benign and not break any mutual
exclusion; furthermore, both the seqlock and busylock have the same
scope.
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210309015950.27688-3-dave@stgolabs.net
This brings the behaviour of the uncontended fast path back to roughly
equivalent to simple spinlocks -- a single atomic op with lock hint.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Acked-by: Waiman Long <longman@redhat.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200724131423.1362108-6-npiggin@gmail.com
This implements the generic paravirt qspinlocks using H_PROD and
H_CONFER to kick and wait.
This uses an un-directed yield to any CPU rather than the directed
yield to a pre-empted lock holder that paravirtualised simple
spinlocks use, that requires no kick hcall. This is something that
could be investigated and improved in future.
Performance results can be found in the commit which added queued
spinlocks.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Waiman Long <longman@redhat.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200724131423.1362108-5-npiggin@gmail.com
These have shown significantly improved performance and fairness when
spinlock contention is moderate to high on very large systems.
With this series including subsequent patches, on a 16 socket 1536
thread POWER9, a stress test such as same-file open/close from all
CPUs gets big speedups, 11620op/s aggregate with simple spinlocks vs
384158op/s (33x faster), where the difference in throughput between
the fastest and slowest thread goes from 7x to 1.4x.
Thanks to the fast path being identical in terms of atomics and
barriers (after a subsequent optimisation patch), single threaded
performance is not changed (no measurable difference).
On smaller systems, performance and fairness seems to be generally
improved. Using dbench on tmpfs as a test (that starts to run into
kernel spinlock contention), a 2-socket OpenPOWER POWER9 system was
tested with bare metal and KVM guest configurations. Results can be
found here:
https://github.com/linuxppc/issues/issues/305#issuecomment-663487453
Observations are:
- Queued spinlocks are equal when contention is insignificant, as
expected and as measured with microbenchmarks.
- When there is contention, on bare metal queued spinlocks have better
throughput and max latency at all points.
- When virtualised, queued spinlocks are slightly worse approaching
peak throughput, but significantly better throughput and max latency
at all points beyond peak, until queued spinlock maximum latency
rises when clients are 2x vCPUs.
The regressions haven't been analysed very well yet, there are a lot
of things that can be tuned, particularly the paravirtualised locking,
but the numbers already look like a good net win even on relatively
small systems.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Waiman Long <longman@redhat.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200724131423.1362108-4-npiggin@gmail.com
