On Mon, 2023-07-10 at 14:57 -0400, Chuck Lever wrote:
> On Mon, Jul 10, 2023 at 02:28:56PM -0400, Jeff Layton wrote:
> > On Mon, 2023-07-10 at 12:42 -0400, Chuck Lever wrote:
> > > From: Chuck Lever <chuck.lever@xxxxxxxxxx>
> > >
> > > I've noticed that client-observed server request latency goes up
> > > simply when the nfsd thread count is increased.
> > >
> > > Walking the whole set of pool threads is memory-inefficient. On a
> > > busy server with many threads, enqueuing a transport will visit all
> > > the threads in the pool quite frequently. This also pulls in the
> > > cache lines for some hot fields in each svc_rqst (namely, rq_flags).
> > >
> > > The svc_xprt_enqueue() call that concerns me most is the one in
> > > svc_rdma_wc_receive(), which is single-threaded per CQ. Slowing
> > > down completion handling limits the total throughput per RDMA
> > > connection.
> > >
> > > Instead, set up a busy bitmap and use find_next_clear_bit, which
> > > should work the same way as RQ_BUSY but will touch only the cache
> > > lines that the bitmap is in. Stick with atomic bit operations to
> > > avoid taking a spinlock during the search.
> > >
> > > Signed-off-by: Chuck Lever <chuck.lever@xxxxxxxxxx>
> > > ---
> > > include/linux/sunrpc/svc.h | 6 ++++--
> > > include/trace/events/sunrpc.h | 1 -
> > > net/sunrpc/svc.c | 24 +++++++++++++++++++-----
> > > net/sunrpc/svc_xprt.c | 26 ++++++++++++++++++++------
> > > 4 files changed, 43 insertions(+), 14 deletions(-)
> > >
> > > diff --git a/include/linux/sunrpc/svc.h b/include/linux/sunrpc/svc.h
> > > index 86377506a514..6669f3eb9ed4 100644
> > > --- a/include/linux/sunrpc/svc.h
> > > +++ b/include/linux/sunrpc/svc.h
> > > @@ -35,6 +35,7 @@ struct svc_pool {
> > > spinlock_t sp_lock; /* protects sp_sockets */
> > > struct list_head sp_sockets; /* pending sockets */
> > > unsigned int sp_nrthreads; /* # of threads in pool */
> > > + unsigned long *sp_busy_map; /* running threads */
> > > struct xarray sp_thread_xa;
> > >
> > > /* statistics on pool operation */
> > > @@ -191,6 +192,8 @@ extern u32 svc_max_payload(const struct svc_rqst *rqstp);
> > > #define RPCSVC_MAXPAGES ((RPCSVC_MAXPAYLOAD+PAGE_SIZE-1)/PAGE_SIZE \
> > > + 2 + 1)
> > >
> > > +#define RPCSVC_MAXPOOLTHREADS (4096)
> > > +
> > > /*
> > > * The context of a single thread, including the request currently being
> > > * processed.
> > > @@ -240,8 +243,7 @@ struct svc_rqst {
> > > #define RQ_SPLICE_OK (4) /* turned off in gss privacy
> > > * to prevent encrypting page
> > > * cache pages */
> > > -#define RQ_BUSY (5) /* request is busy */
> > > -#define RQ_DATA (6) /* request has data */
> > > +#define RQ_DATA (5) /* request has data */
> > > unsigned long rq_flags; /* flags field */
> > > u32 rq_thread_id; /* xarray index */
> > > ktime_t rq_qtime; /* enqueue time */
> > > diff --git a/include/trace/events/sunrpc.h b/include/trace/events/sunrpc.h
> > > index ea43c6059bdb..c07824a254bf 100644
> > > --- a/include/trace/events/sunrpc.h
> > > +++ b/include/trace/events/sunrpc.h
> > > @@ -1676,7 +1676,6 @@ DEFINE_SVCXDRBUF_EVENT(sendto);
> > > svc_rqst_flag(USEDEFERRAL) \
> > > svc_rqst_flag(DROPME) \
> > > svc_rqst_flag(SPLICE_OK) \
> > > - svc_rqst_flag(BUSY) \
> > > svc_rqst_flag_end(DATA)
> > >
> > > #undef svc_rqst_flag
> > > diff --git a/net/sunrpc/svc.c b/net/sunrpc/svc.c
> > > index 109d7f047385..f6305b66fd28 100644
> > > --- a/net/sunrpc/svc.c
> > > +++ b/net/sunrpc/svc.c
> > > @@ -509,6 +509,12 @@ __svc_create(struct svc_program *prog, unsigned int bufsize, int npools,
> > > INIT_LIST_HEAD(&pool->sp_sockets);
> > > spin_lock_init(&pool->sp_lock);
> > > xa_init_flags(&pool->sp_thread_xa, XA_FLAGS_ALLOC);
> > > + pool->sp_busy_map =
> > > + bitmap_alloc_node(RPCSVC_MAXPOOLTHREADS, GFP_KERNEL,
> > > + svc_pool_map_get_node(i));
> > > + if (!pool->sp_busy_map)
> > > + return NULL;
> > > + bitmap_fill(pool->sp_busy_map, RPCSVC_MAXPOOLTHREADS);
> > >
> > > percpu_counter_init(&pool->sp_messages_arrived, 0, GFP_KERNEL);
> > > percpu_counter_init(&pool->sp_sockets_queued, 0, GFP_KERNEL);
> > > @@ -598,6 +604,8 @@ svc_destroy(struct kref *ref)
> > > percpu_counter_destroy(&pool->sp_threads_no_work);
> > >
> > > xa_destroy(&pool->sp_thread_xa);
> > > + bitmap_free(pool->sp_busy_map);
> > > + pool->sp_busy_map = NULL;
> > > }
> > > kfree(serv->sv_pools);
> > > kfree(serv);
> > > @@ -649,7 +657,6 @@ svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
> > >
> > > folio_batch_init(&rqstp->rq_fbatch);
> > >
> > > - __set_bit(RQ_BUSY, &rqstp->rq_flags);
> > > rqstp->rq_server = serv;
> > > rqstp->rq_pool = pool;
> > >
> > > @@ -679,7 +686,7 @@ static struct svc_rqst *
> > > svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
> > > {
> > > struct xa_limit limit = {
> > > - .max = U32_MAX,
> > > + .max = RPCSVC_MAXPOOLTHREADS,
> > > };
> > > struct svc_rqst *rqstp;
> > > int ret;
> > > @@ -724,12 +731,19 @@ struct svc_rqst *svc_pool_wake_idle_thread(struct svc_serv *serv,
> > > struct svc_pool *pool)
> > > {
> > > struct svc_rqst *rqstp;
> > > - unsigned long index;
> > > + unsigned long bit;
> > >
> > > - xa_for_each(&pool->sp_thread_xa, index, rqstp) {
> > > - if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
> > > + /* Check the pool's idle bitmap locklessly so that multiple
> > > + * idle searches can proceed concurrently.
> > > + */
> > > + for_each_clear_bit(bit, pool->sp_busy_map, pool->sp_nrthreads) {
> > > + if (test_and_set_bit(bit, pool->sp_busy_map))
> > > continue;
> > >
> > >
> >
> > I wonder if we also need to set the sp_busy_map bit when we're trying to
> > shrink the threadpool? It seems like the code above that is trying to
> > find a thread to do work could race with the task being killed, such
> > that we try to take up a task that is no longer functional.
>
> What I /think/ will happen is that while svc_pool_victim() is
> holding the xa_lock, it will remove the rqst from the xarray.
> Then when svc_pool_wake_idle_thread() calls xa_load(), it should
> get a NULL, and treat that like a pool that doesn't have an
> available thread to wake.
>
Right, but we're searching under RCU, so a search could find the thing
just before it gets removed. IOW, something like this is possible:
receiving task rpc.nfsd nfsd thread
---------------------------------------------------------------------------------------
svc_pool_wake_idle_thread
xa_load
svc_set_num_threads
svc_stop_kthreads
(remove from xarray)
svc_exit_thread
test_and_set_bit
wake_up_process
Granted, I think this is a potential race in the existing code too, so
this patch probably doesn't make anything worse.
>
> > > + rqstp = xa_load(&pool->sp_thread_xa, bit);
> > > + if (!rqstp)
> > > + break;
> > > +
> > > WRITE_ONCE(rqstp->rq_qtime, ktime_get());
> > > wake_up_process(rqstp->rq_task);
> > > percpu_counter_inc(&pool->sp_threads_woken);
> > > diff --git a/net/sunrpc/svc_xprt.c b/net/sunrpc/svc_xprt.c
> > > index db40f771b60a..f9c9babe0cba 100644
> > > --- a/net/sunrpc/svc_xprt.c
> > > +++ b/net/sunrpc/svc_xprt.c
> > > @@ -735,6 +735,21 @@ rqst_should_sleep(struct svc_rqst *rqstp)
> > > return true;
> > > }
> > >
> > > +static void svc_pool_thread_mark_idle(struct svc_pool *pool,
> > > + struct svc_rqst *rqstp)
> > > +{
> > > + clear_bit_unlock(rqstp->rq_thread_id, pool->sp_busy_map);
> > > +}
> > > +
> > > +/*
> > > + * Note: If we were awoken, then this rqstp has already been marked busy.
> > > + */
> > > +static void svc_pool_thread_mark_busy(struct svc_pool *pool,
> > > + struct svc_rqst *rqstp)
> > > +{
> > > + test_and_set_bit_lock(rqstp->rq_thread_id, pool->sp_busy_map);
> > > +}
> > > +
> > > static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
> > > {
> > > struct svc_pool *pool = rqstp->rq_pool;
> > > @@ -756,18 +771,17 @@ static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
> > > set_current_state(TASK_INTERRUPTIBLE);
> > > smp_mb__before_atomic();
> > > clear_bit(SP_CONGESTED, &pool->sp_flags);
> > > - clear_bit(RQ_BUSY, &rqstp->rq_flags);
> > > - smp_mb__after_atomic();
> > >
> > > - if (likely(rqst_should_sleep(rqstp)))
> > > + if (likely(rqst_should_sleep(rqstp))) {
> > > + svc_pool_thread_mark_idle(pool, rqstp);
> > > time_left = schedule_timeout(timeout);
> > > - else
> > > + } else
> > > __set_current_state(TASK_RUNNING);
> > >
> > > try_to_freeze();
> > >
> > > - set_bit(RQ_BUSY, &rqstp->rq_flags);
> > > - smp_mb__after_atomic();
> > > + svc_pool_thread_mark_busy(pool, rqstp);
> > > +
> > > rqstp->rq_xprt = svc_xprt_dequeue(pool);
> > > if (rqstp->rq_xprt) {
> > > trace_svc_pool_awoken(rqstp);
> > >
> > >
> >
> > --
> > Jeff Layton <jlayton@xxxxxxxxxx>
--
Jeff Layton <jlayton@xxxxxxxxxx>
