lkml.org 
[lkml]   [2020]   [Jan]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    SubjectRe: [PATCH v9 4/5] locking/qspinlock: Introduce starvation avoidance into CNA
    From
    Date
    On 1/14/20 10:59 PM, Alex Kogan wrote:
    > Keep track of the number of intra-node lock handoffs, and force
    > inter-node handoff once this number reaches a preset threshold.
    > The default value for the threshold can be overridden with
    > the new kernel boot command-line option "numa_spinlock_threshold".
    >
    > Signed-off-by: Alex Kogan <alex.kogan@oracle.com>
    > Reviewed-by: Steve Sistare <steven.sistare@oracle.com>
    > Reviewed-by: Waiman Long <longman@redhat.com>
    > ---
    > .../admin-guide/kernel-parameters.txt | 8 ++++
    > kernel/locking/qspinlock.c | 3 ++
    > kernel/locking/qspinlock_cna.h | 41 ++++++++++++++++++-
    > 3 files changed, 51 insertions(+), 1 deletion(-)
    >
    > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
    > index b68cb80e477f..30d79819a3b0 100644
    > --- a/Documentation/admin-guide/kernel-parameters.txt
    > +++ b/Documentation/admin-guide/kernel-parameters.txt
    > @@ -3200,6 +3200,14 @@
    > Not specifying this option is equivalent to
    > numa_spinlock=auto.
    >
    > + numa_spinlock_threshold= [NUMA, PV_OPS]
    > + Set the threshold for the number of intra-node
    > + lock hand-offs before the NUMA-aware spinlock
    > + is forced to be passed to a thread on another NUMA node.
    > + Valid values are in the [0..31] range. Smaller values
    > + result in a more fair, but less performant spinlock, and
    > + vice versa. The default value is 16.
    > +
    > cpu0_hotplug [X86] Turn on CPU0 hotplug feature when
    > CONFIG_BOOTPARAM_HOTPLUG_CPU0 is off.
    > Some features depend on CPU0. Known dependencies are:
    > diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c
    > index 609980a53841..e382d8946ccc 100644
    > --- a/kernel/locking/qspinlock.c
    > +++ b/kernel/locking/qspinlock.c
    > @@ -597,6 +597,9 @@ EXPORT_SYMBOL(queued_spin_lock_slowpath);
    > #if !defined(_GEN_CNA_LOCK_SLOWPATH) && defined(CONFIG_NUMA_AWARE_SPINLOCKS)
    > #define _GEN_CNA_LOCK_SLOWPATH
    >
    > +#undef pv_init_node
    > +#define pv_init_node cna_init_node
    > +
    > #undef pv_wait_head_or_lock
    > #define pv_wait_head_or_lock cna_pre_scan
    >
    > diff --git a/kernel/locking/qspinlock_cna.h b/kernel/locking/qspinlock_cna.h
    > index 8000231f3d51..a2b65f87e6f8 100644
    > --- a/kernel/locking/qspinlock_cna.h
    > +++ b/kernel/locking/qspinlock_cna.h
    > @@ -51,13 +51,25 @@ struct cna_node {
    > int numa_node;
    > u32 encoded_tail;
    > u32 pre_scan_result; /* encoded tail or enum val */
    > + u32 intra_count;
    > };
    >
    > enum {
    > LOCAL_WAITER_FOUND = 2, /* 0 and 1 are reserved for @locked */
    > + FLUSH_SECONDARY_QUEUE = 3,
    > MIN_ENCODED_TAIL
    > };
    >
    > +/*
    > + * Controls the threshold for the number of intra-node lock hand-offs before
    > + * the NUMA-aware variant of spinlock is forced to be passed to a thread on
    > + * another NUMA node. By default, the chosen value provides reasonable
    > + * long-term fairness without sacrificing performance compared to a lock
    > + * that does not have any fairness guarantees. The default setting can
    > + * be changed with the "numa_spinlock_threshold" boot option.
    > + */
    > +unsigned int intra_node_handoff_threshold __ro_after_init = 1 << 16;
    > +
    > static void __init cna_init_nodes_per_cpu(unsigned int cpu)
    > {
    > struct mcs_spinlock *base = per_cpu_ptr(&qnodes[0].mcs, cpu);
    > @@ -97,6 +109,11 @@ static int __init cna_init_nodes(void)
    > }
    > early_initcall(cna_init_nodes);
    >
    > +static __always_inline void cna_init_node(struct mcs_spinlock *node)
    > +{
    > + ((struct cna_node *)node)->intra_count = 0;
    > +}
    > +
    > /* this function is called only when the primary queue is empty */
    > static inline bool cna_try_change_tail(struct qspinlock *lock, u32 val,
    > struct mcs_spinlock *node)
    > @@ -232,7 +249,9 @@ __always_inline u32 cna_pre_scan(struct qspinlock *lock,
    > {
    > struct cna_node *cn = (struct cna_node *)node;
    >
    > - cn->pre_scan_result = cna_scan_main_queue(node, node);
    > + cn->pre_scan_result =
    > + cn->intra_count == intra_node_handoff_threshold ?
    > + FLUSH_SECONDARY_QUEUE : cna_scan_main_queue(node, node);
    >
    > return 0;
    > }
    > @@ -262,6 +281,9 @@ static inline void cna_pass_lock(struct mcs_spinlock *node,
    > * if we acquired the MCS lock when its queue was empty
    > */
    > val = node->locked ? node->locked : 1;
    > + /* inc @intra_count if the secondary queue is not empty */
    > + ((struct cna_node *)next_holder)->intra_count =
    > + cn->intra_count + (node->locked > 1);

    Playing with lock event counts, I would like you to change the meaning
    intra_count parameter that you are tracking. Instead of tracking the
    number of times a lock is passed to a waiter of the same node
    consecutively, I would like you to track the number of times the head
    waiter in the secondary queue has given up its chance to acquire the
    lock because a later waiter has jumped the queue and acquire the lock
    before it. This value determines the worst case latency that a secondary
    queue waiter can experience. So

    @@ -332,8 +334,12 @@ static inline void cna_pass_lock(struct
    mcs_spinlock *node,
                     */
                    val = node->locked ? node->locked : 1;
     
    -               /* inc @intra_count if the secondary queue is not empty */
    -               next_cn->intra_count = cn->intra_count + (node->locked > 1);
    +               /*
    +                * inc @intra_count and pass it down if the secondary queue
    +                * is not empty
    +                */
    +               if (node->locked > 1)
    +                       next_cn->intra_count = cn->intra_count + 1;
            } else if (node->locked > 1) {    /* if secondary queue is not
    empty */
                    /* next holder will be the first node in the secondary
    queue */

    Maybe rename it to jump_count or some other more meaningful name. With
    that change, we could probably reduce the default threshold from 64k to
    maybe 256 or 512.

    I changed the threshold to 256 and run a 96-thread locking stress test
    for 10s, the lock event counts:

    cna_flush_queue=15687
    cna_intra_max=256
    cna_mainscan_hit=13
    cna_merge_queue=15691
    cna_prescan_hit=4344037
    cna_prescan_miss=21
    cna_splice_new=15701
    cna_splice_old=1289
    lock_pending=4384
    lock_slowpath=47998292
    lock_use_node2=16778

    Of the prescan hits, only about 0.4% of that resulted in a queue flush
    which I thought is reasonable. I didn't see any noticeable degradation
    in the performance of the stress test by reducing the threshold from 64k
    to 256.

    Cheers,
    Longman

    \
     
     \ /
      Last update: 2020-01-23 20:56    [W:4.249 / U:0.004 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site