The qspinlock slowpath must ensure that the MCS node is fully initialised
before it can be reached by another other CPU. This is currently enforced
by using a RELEASE operation when updating the tail and also when linking
the node into the waitqueue, since the control dependency off xchg_tail
is insufficient to enforce sufficient ordering, see:
95bcade33a ("locking/qspinlock: Ensure node is initialised before updating prev->next")
Back-to-back RELEASE operations may be expensive on some architectures,
particularly those that implement them using fences under the hood. We
can replace the two RELEASE operations with a single smp_wmb() fence and
use RELAXED operations for the subsequent publishing of the node.
Signed-off-by: Will Deacon <will.deacon@arm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Waiman Long <longman@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: boqun.feng@gmail.com
Cc: linux-arm-kernel@lists.infradead.org
Cc: paulmck@linux.vnet.ibm.com
Link: http://lkml.kernel.org/r/1524738868-31318-12-git-send-email-will.deacon@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The qspinlock locking slowpath utilises a "pending" bit as a simple form
of an embedded test-and-set lock that can avoid the overhead of explicit
queuing in cases where the lock is held but uncontended. This bit is
managed using a cmpxchg() loop which tries to transition the uncontended
lock word from (0,0,0) -> (0,0,1) or (0,0,1) -> (0,1,1).
Unfortunately, the cmpxchg() loop is unbounded and lockers can be starved
indefinitely if the lock word is seen to oscillate between unlocked
(0,0,0) and locked (0,0,1). This could happen if concurrent lockers are
able to take the lock in the cmpxchg() loop without queuing and pass it
around amongst themselves.
This patch fixes the problem by unconditionally setting _Q_PENDING_VAL
using atomic_fetch_or, and then inspecting the old value to see whether
we need to spin on the current lock owner, or whether we now effectively
hold the lock. The tricky scenario is when concurrent lockers end up
queuing on the lock and the lock becomes available, causing us to see
a lockword of (n,0,0). With pending now set, simply queuing could lead
to deadlock as the head of the queue may not have observed the pending
flag being cleared. Conversely, if the head of the queue did observe
pending being cleared, then it could transition the lock from (n,0,0) ->
(0,0,1) meaning that any attempt to "undo" our setting of the pending
bit could race with a concurrent locker trying to set it.
We handle this race by preserving the pending bit when taking the lock
after reaching the head of the queue and leaving the tail entry intact
if we saw pending set, because we know that the tail is going to be
updated shortly.
Signed-off-by: Will Deacon <will.deacon@arm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Waiman Long <longman@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: boqun.feng@gmail.com
Cc: linux-arm-kernel@lists.infradead.org
Cc: paulmck@linux.vnet.ibm.com
Link: http://lkml.kernel.org/r/1524738868-31318-6-git-send-email-will.deacon@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
If a locker taking the qspinlock slowpath reads a lock value indicating
that only the pending bit is set, then it will spin whilst the
concurrent pending->locked transition takes effect.
Unfortunately, there is no guarantee that such a transition will ever be
observed since concurrent lockers could continuously set pending and
hand over the lock amongst themselves, leading to starvation. Whilst
this would probably resolve in practice, it means that it is not
possible to prove liveness properties about the lock and means that lock
acquisition time is unbounded.
Rather than removing the pending->locked spinning from the slowpath
altogether (which has been shown to heavily penalise a 2-threaded
locking stress test on x86), this patch replaces the explicit spinning
with a call to atomic_cond_read_relaxed and allows the architecture to
provide a bound on the number of spins. For architectures that can
respond to changes in cacheline state in their smp_cond_load implementation,
it should be sufficient to use the default bound of 1.
Suggested-by: Waiman Long <longman@redhat.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Waiman Long <longman@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: boqun.feng@gmail.com
Cc: linux-arm-kernel@lists.infradead.org
Cc: paulmck@linux.vnet.ibm.com
Link: http://lkml.kernel.org/r/1524738868-31318-4-git-send-email-will.deacon@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When queuing on the qspinlock, the count field for the current CPU's head
node is incremented. This needn't be atomic because locking in e.g. IRQ
context is balanced and so an IRQ will return with node->count as it
found it.
However, the compiler could in theory reorder the initialisation of
node[idx] before the increment of the head node->count, causing an
IRQ to overwrite the initialised node and potentially corrupt the lock
state.
Avoid the potential for this harmful compiler reordering by placing a
barrier() between the increment of the head node->count and the subsequent
node initialisation.
Signed-off-by: Will Deacon <will.deacon@arm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1518528177-19169-3-git-send-email-will.deacon@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
When a locker ends up queuing on the qspinlock locking slowpath, we
initialise the relevant mcs node and publish it indirectly by updating
the tail portion of the lock word using xchg_tail. If we find that there
was a pre-existing locker in the queue, we subsequently update their
->next field to point at our node so that we are notified when it's our
turn to take the lock.
This can be roughly illustrated as follows:
/* Initialise the fields in node and encode a pointer to node in tail */
tail = initialise_node(node);
/*
* Exchange tail into the lockword using an atomic read-modify-write
* operation with release semantics
*/
old = xchg_tail(lock, tail);
/* If there was a pre-existing waiter ... */
if (old & _Q_TAIL_MASK) {
prev = decode_tail(old);
smp_read_barrier_depends();
/* ... then update their ->next field to point to node.
WRITE_ONCE(prev->next, node);
}
The conditional update of prev->next therefore relies on the address
dependency from the result of xchg_tail ensuring order against the
prior initialisation of node. However, since the release semantics of
the xchg_tail operation apply only to the write portion of the RmW,
then this ordering is not guaranteed and it is possible for the CPU
to return old before the writes to node have been published, consequently
allowing us to point prev->next to an uninitialised node.
This patch fixes the problem by making the update of prev->next a RELEASE
operation, which also removes the reliance on dependency ordering.
Signed-off-by: Will Deacon <will.deacon@arm.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1518528177-19169-2-git-send-email-will.deacon@arm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Queued spinlocks are not used by DEC Alpha, and furthermore operations
such as READ_ONCE() and release/relaxed RMW atomics are being changed
to imply smp_read_barrier_depends(). This commit therefore removes the
now-redundant smp_read_barrier_depends() from queued_spin_lock_slowpath(),
and adjusts the comments accordingly.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>
While this prior commit:
54cf809b95 ("locking,qspinlock: Fix spin_is_locked() and spin_unlock_wait()")
... fixes spin_is_locked() and spin_unlock_wait() for the usage
in ipc/sem and netfilter, it does not in fact work right for the
usage in task_work and futex.
So while the 2 locks crossed problem:
spin_lock(A) spin_lock(B)
if (!spin_is_locked(B)) spin_unlock_wait(A)
foo() foo();
... works with the smp_mb() injected by both spin_is_locked() and
spin_unlock_wait(), this is not sufficient for:
flag = 1;
smp_mb(); spin_lock()
spin_unlock_wait() if (!flag)
// add to lockless list
// iterate lockless list
... because in this scenario, the store from spin_lock() can be delayed
past the load of flag, uncrossing the variables and loosing the
guarantee.
This patch reworks spin_is_locked() and spin_unlock_wait() to work in
both cases by exploiting the observation that while the lock byte
store can be delayed, the contender must have registered itself
visibly in other state contained in the word.
It also allows for architectures to override both functions, as PPC
and ARM64 have an additional issue for which we currently have no
generic solution.
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Giovanni Gherdovich <ggherdovich@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Pan Xinhui <xinhui.pan@linux.vnet.ibm.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Waiman Long <waiman.long@hpe.com>
Cc: Will Deacon <will.deacon@arm.com>
Cc: stable@vger.kernel.org # v4.2 and later
Fixes: 54cf809b95 ("locking,qspinlock: Fix spin_is_locked() and spin_unlock_wait()")
Signed-off-by: Ingo Molnar <mingo@kernel.org>
In an overcommitted guest where some vCPUs have to be halted to make
forward progress in other areas, it is highly likely that a vCPU later
in the spinlock queue will be spinning while the ones earlier in the
queue would have been halted. The spinning in the later vCPUs is then
just a waste of precious CPU cycles because they are not going to
get the lock soon as the earlier ones have to be woken up and take
their turn to get the lock.
This patch implements an adaptive spinning mechanism where the vCPU
will call pv_wait() if the previous vCPU is not running.
Linux kernel builds were run in KVM guest on an 8-socket, 4
cores/socket Westmere-EX system and a 4-socket, 8 cores/socket
Haswell-EX system. Both systems are configured to have 32 physical
CPUs. The kernel build times before and after the patch were:
Westmere Haswell
Patch 32 vCPUs 48 vCPUs 32 vCPUs 48 vCPUs
----- -------- -------- -------- --------
Before patch 3m02.3s 5m00.2s 1m43.7s 3m03.5s
After patch 3m03.0s 4m37.5s 1m43.0s 2m47.2s
For 32 vCPUs, this patch doesn't cause any noticeable change in
performance. For 48 vCPUs (over-committed), there is about 8%
performance improvement.
Signed-off-by: Waiman Long <Waiman.Long@hpe.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Douglas Hatch <doug.hatch@hpe.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Scott J Norton <scott.norton@hpe.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1447114167-47185-8-git-send-email-Waiman.Long@hpe.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch allows one attempt for the lock waiter to steal the lock
when entering the PV slowpath. To prevent lock starvation, the pending
bit will be set by the queue head vCPU when it is in the active lock
spinning loop to disable any lock stealing attempt. This helps to
reduce the performance penalty caused by lock waiter preemption while
not having much of the downsides of a real unfair lock.
The pv_wait_head() function was renamed as pv_wait_head_or_lock()
as it was modified to acquire the lock before returning. This is
necessary because of possible lock stealing attempts from other tasks.
Linux kernel builds were run in KVM guest on an 8-socket, 4
cores/socket Westmere-EX system and a 4-socket, 8 cores/socket
Haswell-EX system. Both systems are configured to have 32 physical
CPUs. The kernel build times before and after the patch were:
Westmere Haswell
Patch 32 vCPUs 48 vCPUs 32 vCPUs 48 vCPUs
----- -------- -------- -------- --------
Before patch 3m15.6s 10m56.1s 1m44.1s 5m29.1s
After patch 3m02.3s 5m00.2s 1m43.7s 3m03.5s
For the overcommited case (48 vCPUs), this patch is able to reduce
kernel build time by more than 54% for Westmere and 44% for Haswell.
Signed-off-by: Waiman Long <Waiman.Long@hpe.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Douglas Hatch <doug.hatch@hpe.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Scott J Norton <scott.norton@hpe.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1447190336-53317-1-git-send-email-Waiman.Long@hpe.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
A queue head CPU, after acquiring the lock, will have to notify
the next CPU in the wait queue that it has became the new queue
head. This involves loading a new cacheline from the MCS node of the
next CPU. That operation can be expensive and add to the latency of
locking operation.
This patch addes code to optmistically prefetch the next MCS node
cacheline if the next pointer is defined and it has been spinning
for the MCS lock for a while. This reduces the locking latency and
improves the system throughput.
The performance change will depend on whether the prefetch overhead
can be hidden within the latency of the lock spin loop. On really
short critical section, there may not be performance gain at all. With
longer critical section, however, it was found to have a performance
boost of 5-10% over a range of different queue depths with a spinlock
loop microbenchmark.
Signed-off-by: Waiman Long <Waiman.Long@hpe.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Douglas Hatch <doug.hatch@hpe.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Scott J Norton <scott.norton@hpe.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1447114167-47185-3-git-send-email-Waiman.Long@hpe.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
For an over-committed guest with more vCPUs than physical CPUs
available, it is possible that a vCPU may be kicked twice before
getting the lock - once before it becomes queue head and once again
before it gets the lock. All these CPU kicking and halting (VMEXIT)
can be expensive and slow down system performance.
This patch adds a new vCPU state (vcpu_hashed) which enables the code
to delay CPU kicking until at unlock time. Once this state is set,
the new lock holder will set _Q_SLOW_VAL and fill in the hash table
on behalf of the halted queue head vCPU. The original vcpu_halted
state will be used by pv_wait_node() only to differentiate other
queue nodes from the qeue head.
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Douglas Hatch <doug.hatch@hp.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Scott J Norton <scott.norton@hp.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1436647018-49734-2-git-send-email-Waiman.Long@hp.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Currently, atomic_cmpxchg() is used to get the lock. However, this
is not really necessary if there is more than one task in the queue
and the queue head don't need to reset the tail code. For that case,
a simple write to set the lock bit is enough as the queue head will
be the only one eligible to get the lock as long as it checks that
both the lock and pending bits are not set. The current pending bit
waiting code will ensure that the bit will not be set as soon as the
tail code in the lock is set.
With that change, the are some slight improvement in the performance
of the queued spinlock in the 5M loop micro-benchmark run on a 4-socket
Westere-EX machine as shown in the tables below.
[Standalone/Embedded - same node]
# of tasks Before patch After patch %Change
---------- ----------- ---------- -------
3 2324/2321 2248/2265 -3%/-2%
4 2890/2896 2819/2831 -2%/-2%
5 3611/3595 3522/3512 -2%/-2%
6 4281/4276 4173/4160 -3%/-3%
7 5018/5001 4875/4861 -3%/-3%
8 5759/5750 5563/5568 -3%/-3%
[Standalone/Embedded - different nodes]
# of tasks Before patch After patch %Change
---------- ----------- ---------- -------
3 12242/12237 12087/12093 -1%/-1%
4 10688/10696 10507/10521 -2%/-2%
It was also found that this change produced a much bigger performance
improvement in the newer IvyBridge-EX chip and was essentially to close
the performance gap between the ticket spinlock and queued spinlock.
The disk workload of the AIM7 benchmark was run on a 4-socket
Westmere-EX machine with both ext4 and xfs RAM disks at 3000 users
on a 3.14 based kernel. The results of the test runs were:
AIM7 XFS Disk Test
kernel JPM Real Time Sys Time Usr Time
----- --- --------- -------- --------
ticketlock 5678233 3.17 96.61 5.81
qspinlock 5750799 3.13 94.83 5.97
AIM7 EXT4 Disk Test
kernel JPM Real Time Sys Time Usr Time
----- --- --------- -------- --------
ticketlock 1114551 16.15 509.72 7.11
qspinlock 2184466 8.24 232.99 6.01
The ext4 filesystem run had a much higher spinlock contention than
the xfs filesystem run.
The "ebizzy -m" test was also run with the following results:
kernel records/s Real Time Sys Time Usr Time
----- --------- --------- -------- --------
ticketlock 2075 10.00 216.35 3.49
qspinlock 3023 10.00 198.20 4.80
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Daniel J Blueman <daniel@numascale.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Douglas Hatch <doug.hatch@hp.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Paolo Bonzini <paolo.bonzini@gmail.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Scott J Norton <scott.norton@hp.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: virtualization@lists.linux-foundation.org
Cc: xen-devel@lists.xenproject.org
Link: http://lkml.kernel.org/r/1429901803-29771-7-git-send-email-Waiman.Long@hp.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch introduces a new generic queued spinlock implementation that
can serve as an alternative to the default ticket spinlock. Compared
with the ticket spinlock, this queued spinlock should be almost as fair
as the ticket spinlock. It has about the same speed in single-thread
and it can be much faster in high contention situations especially when
the spinlock is embedded within the data structure to be protected.
Only in light to moderate contention where the average queue depth
is around 1-3 will this queued spinlock be potentially a bit slower
due to the higher slowpath overhead.
This queued spinlock is especially suit to NUMA machines with a large
number of cores as the chance of spinlock contention is much higher
in those machines. The cost of contention is also higher because of
slower inter-node memory traffic.
Due to the fact that spinlocks are acquired with preemption disabled,
the process will not be migrated to another CPU while it is trying
to get a spinlock. Ignoring interrupt handling, a CPU can only be
contending in one spinlock at any one time. Counting soft IRQ, hard
IRQ and NMI, a CPU can only have a maximum of 4 concurrent lock waiting
activities. By allocating a set of per-cpu queue nodes and used them
to form a waiting queue, we can encode the queue node address into a
much smaller 24-bit size (including CPU number and queue node index)
leaving one byte for the lock.
Please note that the queue node is only needed when waiting for the
lock. Once the lock is acquired, the queue node can be released to
be used later.
Signed-off-by: Waiman Long <Waiman.Long@hp.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Daniel J Blueman <daniel@numascale.com>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Douglas Hatch <doug.hatch@hp.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Paolo Bonzini <paolo.bonzini@gmail.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Scott J Norton <scott.norton@hp.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: virtualization@lists.linux-foundation.org
Cc: xen-devel@lists.xenproject.org
Link: http://lkml.kernel.org/r/1429901803-29771-2-git-send-email-Waiman.Long@hp.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
