> On Jul 3, 2023, at 10:17 PM, NeilBrown <neilb@xxxxxxx> wrote:
>
> On Tue, 04 Jul 2023, Chuck Lever wrote:
>> On Tue, Jul 04, 2023 at 11:26:22AM +1000, NeilBrown wrote:
>>> On Tue, 04 Jul 2023, 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.
>>>>
>>>> List walking is known to be memory-inefficient. On a busy server
>>>> with many threads, enqueuing a transport will walk the "all threads"
>>>> list quite frequently. This also pulls in the cache lines for some
>>>> hot fields in each svc_rqst (namely, rq_flags).
>>>
>>> I think this text could usefully be re-written. By this point in the
>>> series we aren't list walking.
>>>
>>> I'd also be curious to know what latency different you get for just this
>>> change.
>>
>> Not much of a latency difference at lower thread counts.
>>
>> The difference I notice is that with the spinlock version of
>> pool_wake_idle_thread, there is significant lock contention as
>> the thread count increases, and the throughput result of my fio
>> test is lower (outside the result variance).
>>
>>
>>>> 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.
>>>>
>>>> So, avoid walking the "all threads" list to find an idle thread to
>>>> wake. Instead, set up an idle bitmap and use find_next_bit, which
>>>> should work the same way as RQ_BUSY but it will touch only the
>>>> cachelines that the bitmap is in. Stick with atomic bit operations
>>>> to avoid taking the pool lock.
>>>>
>>>> Signed-off-by: Chuck Lever <chuck.lever@xxxxxxxxxx>
>>>> ---
>>>> include/linux/sunrpc/svc.h | 6 ++++--
>>>> include/trace/events/sunrpc.h | 1 -
>>>> net/sunrpc/svc.c | 27 +++++++++++++++++++++------
>>>> net/sunrpc/svc_xprt.c | 30 ++++++++++++++++++++++++------
>>>> 4 files changed, 49 insertions(+), 15 deletions(-)
>>>>
>>>> diff --git a/include/linux/sunrpc/svc.h b/include/linux/sunrpc/svc.h
>>>> index 6f8bfcd44250..27ffcf7371d0 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_idle_map; /* idle threads */
>>>> struct xarray sp_thread_xa;
>>>>
>>>> /* statistics on pool operation */
>>>> @@ -190,6 +191,8 @@ 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.
>>>> @@ -239,8 +242,7 @@ 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 */
>>>
>>> Might this be a good opportunity to convert this to an enum ??
>>>
>>>> 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 ef350f0d8925..d0278e5190ba 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);
>>>> + /* All threads initially marked "busy" */
>>>> + pool->sp_idle_map =
>>>> + bitmap_zalloc_node(RPCSVC_MAXPOOLTHREADS, GFP_KERNEL,
>>>> + svc_pool_map_get_node(i));
>>>> + if (!pool->sp_idle_map)
>>>> + return NULL;
>>>>
>>>> percpu_counter_init(&pool->sp_messages_arrived, 0, GFP_KERNEL);
>>>> percpu_counter_init(&pool->sp_sockets_queued, 0, GFP_KERNEL);
>>>> @@ -596,6 +602,8 @@ svc_destroy(struct kref *ref)
>>>> percpu_counter_destroy(&pool->sp_threads_starved);
>>>>
>>>> xa_destroy(&pool->sp_thread_xa);
>>>> + bitmap_free(pool->sp_idle_map);
>>>> + pool->sp_idle_map = NULL;
>>>> }
>>>> kfree(serv->sv_pools);
>>>> kfree(serv);
>>>> @@ -647,7 +655,6 @@ 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;
>>>>
>>>> @@ -677,7 +684,7 @@ static struct svc_rqst *
>>>> svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
>>>> {
>>>> static const struct xa_limit limit = {
>>>> - .max = U32_MAX,
>>>> + .max = RPCSVC_MAXPOOLTHREADS,
>>>> };
>>>> struct svc_rqst *rqstp;
>>>> int ret;
>>>> @@ -722,12 +729,19 @@ 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_set_bit(bit, pool->sp_idle_map, pool->sp_nrthreads) {
>>>> + if (!test_and_clear_bit(bit, pool->sp_idle_map))
>>>> continue;
>>>
>>> I would really rather the map was "sp_busy_map". (initialised with bitmap_fill())
>>> Then you could "test_and_set_bit_lock()" and later "clear_bit_unlock()"
>>> and so get all the required memory barriers.
>>> What we are doing here is locking a particular thread for a task, so
>>> "lock" is an appropriate description of what is happening.
>>> See also svc_pool_thread_mark_* below.
>>>
>>>>
>>>> + 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);
>>>> @@ -767,7 +781,8 @@ svc_pool_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *stat
>>>> }
>>>>
>>>> found_pool:
>>>> - rqstp = xa_find(&pool->sp_thread_xa, &zero, U32_MAX, XA_PRESENT);
>>>> + rqstp = xa_find(&pool->sp_thread_xa, &zero, RPCSVC_MAXPOOLTHREADS,
>>>> + XA_PRESENT);
>>>> if (rqstp) {
>>>> __xa_erase(&pool->sp_thread_xa, rqstp->rq_thread_id);
>>>> task = rqstp->rq_task;
>>>> diff --git a/net/sunrpc/svc_xprt.c b/net/sunrpc/svc_xprt.c
>>>> index 7709120b45c1..2844b32c16ea 100644
>>>> --- a/net/sunrpc/svc_xprt.c
>>>> +++ b/net/sunrpc/svc_xprt.c
>>>> @@ -735,6 +735,25 @@ rqst_should_sleep(struct svc_rqst *rqstp)
>>>> return true;
>>>> }
>>>>
>>>> +static void svc_pool_thread_mark_idle(struct svc_pool *pool,
>>>> + struct svc_rqst *rqstp)
>>>> +{
>>>> + smp_mb__before_atomic();
>>>> + set_bit(rqstp->rq_thread_id, pool->sp_idle_map);
>>>> + smp_mb__after_atomic();
>>>> +}
>>>
>>> There memory barriers above and below bother me. There is no comment
>>> telling me what they are protecting against.
>>> I would rather svc_pool_thread_mark_idle - which unlocks the thread -
>>> were
>>>
>>> clear_bit_unlock(rqstp->rq_thread_id, pool->sp_busy_map);
>>>
>>> and that svc_pool_thread_mark_busy were
>>>
>>> test_and_set_bit_lock(rqstp->rq_thread_id, pool->sp_busy_map);
>>>
>>> Then it would be more obvious what was happening.
>>
>> Not obvious to me, but that's very likely because I'm not clear what
>> clear_bit_unlock() does. :-)
>
> In general, any "lock" operation (mutex, spin, whatever) is (and must
> be) and "acquire" type operations which imposes a memory barrier so that
> read requests *after* the lock cannot be satisfied with data from
> *before* the lock. The read must access data after the lock.
> Conversely any "unlock" operations is a "release" type operation which
> imposes a memory barrier so that any write request *before* the unlock
> must not be delayed until *after* the unlock. The write must complete
> before the unlock.
>
> This is exactly what you would expect of locking - it creates a closed
> code region that is properly ordered w.r.t comparable closed regions.
>
> test_and_set_bit_lock() and clear_bit_unlock() provide these expected
> semantics for bit operations.
Your explanation is more clear than what I read in Documentation/atomic*
so thanks. I feel a little more armed to make good use of it.
> New code should (almost?) never have explicit memory barriers like
> smp_mb__after_atomic().
> It should use one of the many APIs with _acquire or _release suffixes,
> or with the more explicit _lock or _unlock.
Out of curiosity, is "should never have explicit memory barriers"
documented somewhere? I've been accused of skimming when I read, so
I might have missed it.
--
Chuck Lever
