Re: [PATCH v2 4/4] /proc/stat: Call kstat_irqs_usr() only for active IRQs

[Waiman Long]

On 01/09/2019 12:23 PM, Waiman Long wrote:
> After skipping the percpu summation of non-active IRQs on a 4-socket
> Broadwell system with about 3k IRQs, about half of the CPU cycles were
> spent in the kstat_irqs() call. The majority of which were used to look
> up the IRQ descriptors for the corresponding IRQ numbers.
>
> We can recoup a lot of those lost cycles by calling kstat_irqs_usr()
> only for those IRQs that are active. A bitmap is now used to keep track
> of the list of the active IRQs. Changes in nr_active_irqs count will
> cause the code to rescan all the IRQs and repopulate the bitmap.
>
> On the same 4-socket server, the introduction of this patch further
> reduces the system time of reading /proc/stat 5k times from 8.048s
> to 5.817s. This is a another time reduction of 28%.
>
> Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
> ---
>  fs/proc/stat.c | 84 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 84 insertions(+)
>
> diff --git a/fs/proc/stat.c b/fs/proc/stat.c
> index 4b06f1b..5e2a398 100644
> --- a/fs/proc/stat.c
> +++ b/fs/proc/stat.c
> @@ -93,6 +93,25 @@ static u64 compute_stat_irqs_sum(void)
>  }
>  
>  /*
> + * Write the given number of space separated '0' into the sequence file.
> + */
> +static void write_zeros(struct seq_file *p, int cnt)
> +{
> +	/* String of 16 '0's */
> +	static const char zeros[] = " 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0";
> +
> +	while (cnt > 0) {
> +		if (cnt >= 16) {
> +			seq_write(p, zeros, 32);
> +			cnt -= 16;
> +		} else {
> +			seq_write(p, zeros, 2 * cnt);
> +			cnt = 0;
> +		}
> +	}
> +}
> +
> +/*
>   * Print out the "intr" line of /proc/stat.
>   */
>  static void show_stat_irqs(struct seq_file *p)
> @@ -100,9 +119,74 @@ static void show_stat_irqs(struct seq_file *p)
>  	int i;
>  
>  	seq_put_decimal_ull(p, "intr ", compute_stat_irqs_sum());
> +
> +	if (IS_ENABLED(CONFIG_SMP) && (nr_cpu_ids >= 10) && (nr_irqs >= 256)) {
> +		/*
> +		 * On systems with 10 or more CPUs and 256 or more IRQs,
> +		 * we used a bitmap to keep track of the number of active
> +		 * IRQs and call kstat_irqs_usr() only for those IRQs.
> +		 * The bitmap will be refreshed whenever nr_active_irqs
> +		 * changes.
> +		 */
> +		extern atomic_t nr_active_irqs;
> +		static DEFINE_MUTEX(irqs_mutex);
> +		static int last_irq = -1;
> +		static int bitmap_size, active_irqs;
> +		static unsigned long *bitmap;
> +		int current_irqs = atomic_read(&nr_active_irqs);
> +
> +		mutex_lock(&irqs_mutex);
> +		if (current_irqs != active_irqs) {
> +			/*
> +			 * Rescan all the IRQs for active ones.
> +			 */
> +			if (nr_irqs > bitmap_size) {
> +				static unsigned long *new_bitmap;
> +				static int new_size;
> +
> +				new_size = BITS_TO_LONGS(nr_irqs)*sizeof(long);
> +				new_bitmap = (unsigned long *)krealloc(bitmap,
> +						new_size, GFP_KERNEL);
> +				if (!new_bitmap)
> +					goto fallback;
> +				bitmap = new_bitmap;
> +				bitmap_size = new_size;
> +			}
> +			memset(bitmap, 0, bitmap_size/BITS_PER_BYTE);
> +			last_irq = 0;

Sorry, last_irq should be initialized to -1 here.

Cheers,
Longman