[Devel] [PATCH rh7 3/4] /proc/vz/latency: distinguish atomic allocations in irq from in task atomics.

[Pavel Borzenkov]

> On 17 Aug 2018, at 19:40, Andrey Ryabinin <aryabinin at virtuozzo.com> wrote:
> 
> Add to /proc/vz/latency 'alocirq' allocation type which shows allocation latencies
> done in irq contexts. 'alocatomic' now shows atomic allocations in task contexts.
> Also add 'Per-CPU alloc irq' which shows per-cpu 'alocirq' numbers.

Any chance you can add maximums here as well?

> 
> Example of new output:
> 
> Version: 2.6
> 
> Latencies:
> Type                         Lat            Total_lat                Calls
> scheduling:                    0                    0                    0
> alocatomic:                    0                    0                27615
> aloclow:                 2000000            738000000              3166625
> alochigh:                      0                    0                    0
> aloclowmp:                     0              4000000                15677
> alochighmp:                    0                    0                    0
> alocirq:                 1000000             81000000               292975
> swap_in:                   10710              1553913                   25
> page_in:                   50850           1141057002              2092877
> 
> Averages:
> Type                        Avg1                 Avg5                Avg15
> scheduling:                    0                    0                    0
> alocatomic:                    0                    0                    0
> aloclow:                  558150               313968               165801
> alochigh:                      0                    0                    0
> aloclowmp:                  5535                 9701                 4629
> alochighmp:                    0                    0                    0
> alocirq:                  309600               163409                66832
> swap_in:                   29962                18540                 7244
> page_in:                  177194                60133                29625
> 
> Per-CPU alloc irq:
> Type                         Lat            Total_lat                Calls
> cpu0                           0                    0                    3
> cpu1                           0                    0                    0
> cpu2                           0                    0                    0
> cpu3                     1000000             87000000               316298
> 
> https://jira.sw.ru/browse/PSBM-87797
> Signed-off-by: Andrey Ryabinin <aryabinin at virtuozzo.com>
> Cc: Pavel Borzenkov <Pavel.Borzenkov at acronis.com>
> ---
> include/linux/kstat.h |  1 +
> kernel/ve/vzstat.c    | 21 +++++++++++++++++++--
> mm/page_alloc.c       | 10 +++++++---
> 3 files changed, 27 insertions(+), 5 deletions(-)
> 
> diff --git a/include/linux/kstat.h b/include/linux/kstat.h
> index 4fff25bcb93f..b268752f2e15 100644
> --- a/include/linux/kstat.h
> +++ b/include/linux/kstat.h
> @@ -7,6 +7,7 @@ enum {
> 	KSTAT_ALLOCSTAT_HIGH,
> 	KSTAT_ALLOCSTAT_LOW_MP,
> 	KSTAT_ALLOCSTAT_HIGH_MP,
> +	KSTAT_ALLOCSTAT_IRQ,
> 	KSTAT_ALLOCSTAT_NR,
> };
> 
> diff --git a/kernel/ve/vzstat.c b/kernel/ve/vzstat.c
> index 69cbb38210c0..c8cb0b525aae 100644
> --- a/kernel/ve/vzstat.c
> +++ b/kernel/ve/vzstat.c
> @@ -7,6 +7,7 @@
> 
> #include <linux/sched.h>
> #include <linux/module.h>
> +#include <linux/cpu.h>
> #include <linux/mm.h>
> #include <linux/ve.h>
> #include <linux/ve_proto.h>
> @@ -32,7 +33,8 @@ static const char *alloc_descr[KSTAT_ALLOCSTAT_NR] = {
> 	"aloclow:",
> 	"alochigh:",
> 	"aloclowmp:",
> -	"alochighmp:"
> +	"alochighmp:",
> +	"alocirq:"
> };
> 
> /*
> @@ -175,11 +177,12 @@ static void avglat_seq_show(struct seq_file *m,
> static int latency_seq_show(struct seq_file *m, void *v)
> {
> 	int i;
> +	int cpu;
> 
> 	if (!v)
> 		return 0;
> 
> -	seq_puts(m, "Version: 2.5\n");
> +	seq_puts(m, "Version: 2.6\n");
> 
> 	seq_puts(m, "\nLatencies:\n");
> 	seq_printf(m, "%-11s %20s %20s %20s\n",
> @@ -201,6 +204,20 @@ static int latency_seq_show(struct seq_file *m, void *v)
> 	avglat_seq_show(m, "swap_in:", kstat_glob.swap_in.avg);
> 	avglat_seq_show(m, "page_in:", kstat_glob.page_in.avg);
> 
> +	seq_puts(m, "\nPer-CPU alloc irq:\n");
> +	seq_printf(m, "%-11s %20s %20s %20s\n",
> +			"Type", "Lat", "Total_lat", "Calls");
> +
> +	get_online_cpus();
> +	for_each_online_cpu(cpu) {
> +		struct kstat_lat_pcpu_snap_struct *snap;
> +
> +		snap = per_cpu_ptr(kstat_glob.alloc_lat[KSTAT_ALLOCSTAT_IRQ].cur, cpu);
> +		seq_printf(m, "cpu%-8d %20Lu %20Lu %20lu\n", cpu,
> +			snap->maxlat, snap->totlat, snap->count);
> +	}
> +	put_online_cpus();
> +
> 	return 0;
> }
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 179bb0c351b2..9d8c9e4eb970 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -3222,13 +3222,17 @@ static void __alloc_collect_stats(gfp_t gfp_mask, unsigned int order,
> 	int ind, cpu;
> 
> 	time = jiffies_to_usecs(jiffies - time) * 1000;
> -	if (!(gfp_mask & __GFP_WAIT))
> -		ind = KSTAT_ALLOCSTAT_ATOMIC;
> -	else
> +	if (!(gfp_mask & __GFP_WAIT)) {
> +		if (in_task())
> +			ind = KSTAT_ALLOCSTAT_ATOMIC;
> +		else
> +			ind = KSTAT_ALLOCSTAT_IRQ;
> +	} else {
> 		if (order > 0)
> 			ind = KSTAT_ALLOCSTAT_LOW_MP;
> 		else
> 			ind = KSTAT_ALLOCSTAT_LOW;
> +	}
> 
> 	local_irq_save(flags);
> 	cpu = smp_processor_id();
> -- 
> 2.16.4
>