Re: [PATCH 3/4] block/mq-deadline: Stop using per-CPU counters

[Damien Le Moal <Damien.LeMoal@wdc.com>]

On 2021/09/24 8:27, Bart Van Assche wrote:
> Calculating the sum over all CPUs of per-CPU counters frequently is
> inefficient. Hence switch from per-CPU to individual counters. Three
> counters are protected by the mq-deadline spinlock since these are
> only accessed from contexts that already hold that spinlock. The fourth
> counter is atomic because protecting it with the mq-deadline spinlock
> would trigger lock contention.
> 
> Cc: Damien Le Moal <damien.lemoal@wdc.com>
> Cc: Niklas Cassel <Niklas.Cassel@wdc.com>
> Cc: Hannes Reinecke <hare@suse.de>
> Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> ---
>  block/mq-deadline.c | 124 ++++++++++++++++++++------------------------
>  1 file changed, 56 insertions(+), 68 deletions(-)
> 
> diff --git a/block/mq-deadline.c b/block/mq-deadline.c
> index 6deb4306bcf3..b0cfc84a4e6b 100644
> --- a/block/mq-deadline.c
> +++ b/block/mq-deadline.c
> @@ -51,17 +51,16 @@ enum dd_prio {
>  
>  enum { DD_PRIO_COUNT = 3 };
>  
> -/* I/O statistics per I/O priority. */
> +/*
> + * I/O statistics per I/O priority. It is fine if these counters overflow.
> + * What matters is that these counters are at least as wide as
> + * log2(max_outstanding_requests).
> + */
>  struct io_stats_per_prio {
> -	local_t inserted;
> -	local_t merged;
> -	local_t dispatched;
> -	local_t completed;
> -};
> -
> -/* I/O statistics for all I/O priorities (enum dd_prio). */
> -struct io_stats {
> -	struct io_stats_per_prio stats[DD_PRIO_COUNT];
> +	uint32_t inserted;
> +	uint32_t merged;
> +	uint32_t dispatched;
> +	atomic_t completed;

Why not use 64-bits types (regular unsigned long long and atomic64_t) ?

>  };
>  
>  /*
> @@ -74,6 +73,7 @@ struct dd_per_prio {
>  	struct list_head fifo_list[DD_DIR_COUNT];
>  	/* Next request in FIFO order. Read, write or both are NULL. */
>  	struct request *next_rq[DD_DIR_COUNT];
> +	struct io_stats_per_prio stats;
>  };
>  
>  struct deadline_data {
> @@ -88,8 +88,6 @@ struct deadline_data {
>  	unsigned int batching;		/* number of sequential requests made */
>  	unsigned int starved;		/* times reads have starved writes */
>  
> -	struct io_stats __percpu *stats;
> -
>  	/*
>  	 * settings that change how the i/o scheduler behaves
>  	 */
> @@ -103,33 +101,6 @@ struct deadline_data {
>  	spinlock_t zone_lock;
>  };
>  
> -/* Count one event of type 'event_type' and with I/O priority 'prio' */
> -#define dd_count(dd, event_type, prio) do {				\
> -	struct io_stats *io_stats = get_cpu_ptr((dd)->stats);		\
> -									\
> -	BUILD_BUG_ON(!__same_type((dd), struct deadline_data *));	\
> -	BUILD_BUG_ON(!__same_type((prio), enum dd_prio));		\
> -	local_inc(&io_stats->stats[(prio)].event_type);			\
> -	put_cpu_ptr(io_stats);						\
> -} while (0)
> -
> -/*
> - * Returns the total number of dd_count(dd, event_type, prio) calls across all
> - * CPUs. No locking or barriers since it is fine if the returned sum is slightly
> - * outdated.
> - */
> -#define dd_sum(dd, event_type, prio) ({					\
> -	unsigned int cpu;						\
> -	u32 sum = 0;							\
> -									\
> -	BUILD_BUG_ON(!__same_type((dd), struct deadline_data *));	\
> -	BUILD_BUG_ON(!__same_type((prio), enum dd_prio));		\
> -	for_each_present_cpu(cpu)					\
> -		sum += local_read(&per_cpu_ptr((dd)->stats, cpu)->	\
> -				  stats[(prio)].event_type);		\
> -	sum;								\
> -})
> -
>  /* Maps an I/O priority class to a deadline scheduler priority. */
>  static const enum dd_prio ioprio_class_to_prio[] = {
>  	[IOPRIO_CLASS_NONE]	= DD_BE_PRIO,
> @@ -233,7 +204,9 @@ static void dd_merged_requests(struct request_queue *q, struct request *req,
>  	const u8 ioprio_class = dd_rq_ioclass(next);
>  	const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
>  
> -	dd_count(dd, merged, prio);
> +	lockdep_assert_held(&dd->lock);
> +
> +	dd->per_prio[prio].stats.merged++;
>  
>  	/*
>  	 * if next expires before rq, assign its expire time to rq
> @@ -273,7 +246,11 @@ deadline_move_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
>  /* Number of requests queued for a given priority level. */
>  static u32 dd_queued(struct deadline_data *dd, enum dd_prio prio)
>  {
> -	return dd_sum(dd, inserted, prio) - dd_sum(dd, completed, prio);
> +	const struct io_stats_per_prio *stats = &dd->per_prio[prio].stats;
> +
> +	lockdep_assert_held(&dd->lock);
> +
> +	return stats->inserted - atomic_read(&stats->completed);
>  }
>  
>  /*
> @@ -463,7 +440,7 @@ static struct request *__dd_dispatch_request(struct deadline_data *dd,
>  done:
>  	ioprio_class = dd_rq_ioclass(rq);
>  	prio = ioprio_class_to_prio[ioprio_class];
> -	dd_count(dd, dispatched, prio);
> +	dd->per_prio[prio].stats.dispatched++;
>  	/*
>  	 * If the request needs its target zone locked, do it.
>  	 */
> @@ -542,19 +519,22 @@ static void dd_exit_sched(struct elevator_queue *e)
>  
>  	for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
>  		struct dd_per_prio *per_prio = &dd->per_prio[prio];
> +		const struct io_stats_per_prio *stats = &per_prio->stats;
> +		uint32_t queued;
>  
>  		WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_READ]));
>  		WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_WRITE]));
> -		WARN_ONCE(dd_queued(dd, prio) != 0,
> +
> +		spin_lock(&dd->lock);
> +		queued = dd_queued(dd, prio);
> +		spin_unlock(&dd->lock);
> +
> +		WARN_ONCE(queued,
>  			  "statistics for priority %d: i %u m %u d %u c %u\n",
> -			  prio, dd_sum(dd, inserted, prio),
> -			  dd_sum(dd, merged, prio),
> -			  dd_sum(dd, dispatched, prio),
> -			  dd_sum(dd, completed, prio));
> +			  prio, stats->inserted, stats->merged,
> +			  stats->dispatched, atomic_read(&stats->completed));
>  	}
>  
> -	free_percpu(dd->stats);
> -
>  	kfree(dd);
>  }
>  
> @@ -578,11 +558,6 @@ static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
>  
>  	eq->elevator_data = dd;
>  
> -	dd->stats = alloc_percpu_gfp(typeof(*dd->stats),
> -				     GFP_KERNEL | __GFP_ZERO);
> -	if (!dd->stats)
> -		goto free_dd;
> -
>  	for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
>  		struct dd_per_prio *per_prio = &dd->per_prio[prio];
>  
> @@ -604,9 +579,6 @@ static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
>  	q->elevator = eq;
>  	return 0;
>  
> -free_dd:
> -	kfree(dd);
> -
>  put_eq:
>  	kobject_put(&eq->kobj);
>  	return ret;
> @@ -689,8 +661,9 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
>  	blk_req_zone_write_unlock(rq);
>  
>  	prio = ioprio_class_to_prio[ioprio_class];
> +	per_prio = &dd->per_prio[prio];
>  	if (!rq->elv.priv[0]) {
> -		dd_count(dd, inserted, prio);
> +		per_prio->stats.inserted++;
>  		rq->elv.priv[0] = (void *)(uintptr_t)1;
>  	}
>  
> @@ -701,7 +674,6 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
>  
>  	trace_block_rq_insert(rq);
>  
> -	per_prio = &dd->per_prio[prio];
>  	if (at_head) {
>  		list_add(&rq->queuelist, &per_prio->dispatch);
>  	} else {
> @@ -779,7 +751,7 @@ static void dd_finish_request(struct request *rq)
>  	if (!rq->elv.priv[0])
>  		return;
>  
> -	dd_count(dd, completed, prio);
> +	atomic_inc(&per_prio->stats.completed);
>  
>  	if (blk_queue_is_zoned(q)) {
>  		unsigned long flags;
> @@ -966,28 +938,44 @@ static int dd_queued_show(void *data, struct seq_file *m)
>  {
>  	struct request_queue *q = data;
>  	struct deadline_data *dd = q->elevator->elevator_data;
> +	u32 rt, be, idle;
> +
> +	spin_lock(&dd->lock);
> +	rt = dd_queued(dd, DD_RT_PRIO);
> +	be = dd_queued(dd, DD_BE_PRIO);
> +	idle = dd_queued(dd, DD_IDLE_PRIO);
> +	spin_unlock(&dd->lock);
> +
> +	seq_printf(m, "%u %u %u\n", rt, be, idle);
>  
> -	seq_printf(m, "%u %u %u\n", dd_queued(dd, DD_RT_PRIO),
> -		   dd_queued(dd, DD_BE_PRIO),
> -		   dd_queued(dd, DD_IDLE_PRIO));
>  	return 0;
>  }
>  
>  /* Number of requests owned by the block driver for a given priority. */
>  static u32 dd_owned_by_driver(struct deadline_data *dd, enum dd_prio prio)
>  {
> -	return dd_sum(dd, dispatched, prio) + dd_sum(dd, merged, prio)
> -		- dd_sum(dd, completed, prio);
> +	const struct io_stats_per_prio *stats = &dd->per_prio[prio].stats;
> +
> +	lockdep_assert_held(&dd->lock);
> +
> +	return stats->dispatched + stats->merged -
> +		atomic_read(&stats->completed);
>  }
>  
>  static int dd_owned_by_driver_show(void *data, struct seq_file *m)
>  {
>  	struct request_queue *q = data;
>  	struct deadline_data *dd = q->elevator->elevator_data;
> +	u32 rt, be, idle;
> +
> +	spin_lock(&dd->lock);
> +	rt = dd_owned_by_driver(dd, DD_RT_PRIO);
> +	be = dd_owned_by_driver(dd, DD_BE_PRIO);
> +	idle = dd_owned_by_driver(dd, DD_IDLE_PRIO);
> +	spin_unlock(&dd->lock);
> +
> +	seq_printf(m, "%u %u %u\n", rt, be, idle);
>  
> -	seq_printf(m, "%u %u %u\n", dd_owned_by_driver(dd, DD_RT_PRIO),
> -		   dd_owned_by_driver(dd, DD_BE_PRIO),
> -		   dd_owned_by_driver(dd, DD_IDLE_PRIO));
>  	return 0;
>  }
>  
> 

Apart from the nit/comment above, looks OK to me.

Reviewed-by: Damien Le Moal <damien.lemoal@wdc.com>

-- 
Damien Le Moal
Western Digital Research