Correct mem.vmstat.pgsteal* mem.vmstat.pgscan metrics on RHEL7 and Fedora 25

[wcohen]

PCP on x86_64 RHEL7 and Fedora 25 lists the following pgsteal metrics:

$ pminfo mem.vmstat|grep pgsteal
mem.vmstat.pgsteal_dma
mem.vmstat.pgsteal_dma32
mem.vmstat.pgsteal_high
mem.vmstat.pgsteal_movable
mem.vmstat.pgsteal_normal

However, doing pmval for each of these pgsteal metrics on RHEL7 and F25 return "No values available". Those metrics do not appear to line up with what /proc/vmstat provide for RHEL7 and F25:

RHEL7:

$ more /proc/vmstat |grep pgsteal
pgsteal_kswapd_dma 0
pgsteal_kswapd_dma32 0
pgsteal_kswapd_normal 0
pgsteal_kswapd_movable 0
pgsteal_direct_dma 0
pgsteal_direct_dma32 0
pgsteal_direct_normal 0
pgsteal_direct_movable 0

F25:
$ more /proc/vmstat |grep pgsteal
pgsteal_kswapd 3543984
pgsteal_direct 290756

For Fedora 25 the pgscan metrics don't match with what is available in /proc/vmstat:

[wcohen@santana perf]$ pminfo mem.vmstat|grep pgscan
mem.vmstat.pgscan_direct_dma
mem.vmstat.pgscan_direct_dma32
mem.vmstat.pgscan_direct_high
mem.vmstat.pgscan_direct_movable
mem.vmstat.pgscan_direct_normal
mem.vmstat.pgscan_kswapd_dma
mem.vmstat.pgscan_kswapd_dma32
mem.vmstat.pgscan_kswapd_high
mem.vmstat.pgscan_kswapd_movable
mem.vmstat.pgscan_kswapd_normal
[wcohen@santana perf]$ more /proc/vmstat |grep pgscan
pgscan_kswapd 3679442
pgscan_direct 424948
pgscan_direct_throttle 0

[wcohen]

I am not sure how many of these missing metrics in /proc/vmstat are generally useful in addition to the pgscan and pgsteal ones. Wrote a one line script to see /proc/vmstat metrics listed are missing from PCP mem.vmstat.*:

for i in cat /proc/vmstat|awk '{print $1}'; do pmprobe mem.vmstat.${i}; done | awk '$2 < 0 {print $1}' |sort

On Fedora 25:
$ for i in cat /proc/vmstat|awk '{print $1}'; do pmprobe mem.vmstat.${i}; done | awk '$2 < 0 {print $1}' | sort
mem.vmstat.allocstall_dma
mem.vmstat.allocstall_dma32
mem.vmstat.allocstall_movable
mem.vmstat.allocstall_normal
mem.vmstat.compact_daemon_wake
mem.vmstat.compact_free_scanned
mem.vmstat.compact_isolated
mem.vmstat.compact_migrate_scanned
mem.vmstat.drop_pagecache
mem.vmstat.drop_slab
mem.vmstat.nr_file_pages
mem.vmstat.nr_free_cma
mem.vmstat.nr_pages_scanned
mem.vmstat.nr_shmem_hugepages
mem.vmstat.nr_shmem_pmdmapped
mem.vmstat.nr_vmscan_immediate_reclaim
mem.vmstat.nr_zone_active_anon
mem.vmstat.nr_zone_active_file
mem.vmstat.nr_zone_inactive_anon
mem.vmstat.nr_zone_inactive_file
mem.vmstat.nr_zone_unevictable
mem.vmstat.nr_zone_write_pending
mem.vmstat.nr_zspages
mem.vmstat.numa_hint_faults
mem.vmstat.numa_hint_faults_local
mem.vmstat.numa_huge_pte_updates
mem.vmstat.numa_pages_migrated
mem.vmstat.numa_pte_updates
mem.vmstat.pglazyfreed
mem.vmstat.pgmigrate_fail
mem.vmstat.pgmigrate_success
mem.vmstat.pgrefill
mem.vmstat.pgscan_direct
mem.vmstat.pgscan_direct_throttle
mem.vmstat.pgscan_kswapd
mem.vmstat.pgskip_dma
mem.vmstat.pgskip_dma32
mem.vmstat.pgskip_movable
mem.vmstat.pgskip_normal
mem.vmstat.pgsteal_direct
mem.vmstat.pgsteal_kswapd
mem.vmstat.thp_deferred_split_page
mem.vmstat.thp_file_alloc
mem.vmstat.thp_file_mapped
mem.vmstat.thp_split_page
mem.vmstat.thp_split_page_failed
mem.vmstat.thp_split_pmd
mem.vmstat.workingset_activate
mem.vmstat.workingset_nodereclaim
mem.vmstat.workingset_refault

On RHEL7:

$ for i in cat /proc/vmstat|awk '{print $1}'; do pmprobe mem.vmstat.${i}; done | awk '$2 < 0 {print $1}' |sort
mem.vmstat.compact_free_scanned
mem.vmstat.compact_isolated
mem.vmstat.compact_migrate_scanned
mem.vmstat.drop_pagecache
mem.vmstat.drop_slab
mem.vmstat.nr_alloc_batch
mem.vmstat.nr_file_pages
mem.vmstat.nr_free_cma
mem.vmstat.nr_vmscan_immediate_reclaim
mem.vmstat.numa_hint_faults
mem.vmstat.numa_hint_faults_local
mem.vmstat.numa_huge_pte_updates
mem.vmstat.numa_pages_migrated
mem.vmstat.numa_pte_updates
mem.vmstat.pgmigrate_fail
mem.vmstat.pgmigrate_success
mem.vmstat.pgscan_direct_throttle
mem.vmstat.pgsteal_direct_dma
mem.vmstat.pgsteal_direct_dma32
mem.vmstat.pgsteal_direct_movable
mem.vmstat.pgsteal_direct_normal
mem.vmstat.pgsteal_kswapd_dma
mem.vmstat.pgsteal_kswapd_dma32
mem.vmstat.pgsteal_kswapd_movable
mem.vmstat.pgsteal_kswapd_normal
mem.vmstat.workingset_activate
mem.vmstat.workingset_nodereclaim
mem.vmstat.workingset_refault

[fche]

The way that src/pmdas/linux/proc_vmstat.c is written the metrics are hard-coded in C in the vmstat_fields[] array. When it scans /proc/vmstat, it does a linear search against the vmstat_fields[] list to deposit parsed fields in the global _proc_vmstat_t structure. It's not particularly efficient, nor flexible.

It could instead generate the vmstat pmns entries dynamically from the actual contents of /proc/vmstat, and pass every value out as an int64_t, and ditch all the hard-coded proc_vmstat_t etc. machinery.

[myllynen]

On Fedora 27 with PCP 3.12.2 / kernel 4.13 pgscan/pgsteal metrics are missing completely.

With PCP 3.12.2 we have:

$ pminfo | grep -E '(pgscan|pgsteal)' | sort
mem.vmstat.pgscan_direct_dma
mem.vmstat.pgscan_direct_dma32
mem.vmstat.pgscan_direct_high
mem.vmstat.pgscan_direct_movable
mem.vmstat.pgscan_direct_normal
mem.vmstat.pgscan_kswapd_dma
mem.vmstat.pgscan_kswapd_dma32
mem.vmstat.pgscan_kswapd_high
mem.vmstat.pgscan_kswapd_movable
mem.vmstat.pgscan_kswapd_normal
mem.vmstat.pgsteal_direct_dma
mem.vmstat.pgsteal_direct_dma32
mem.vmstat.pgsteal_direct_movable
mem.vmstat.pgsteal_direct_normal
mem.vmstat.pgsteal_dma
mem.vmstat.pgsteal_dma32
mem.vmstat.pgsteal_high
mem.vmstat.pgsteal_kswapd_dma
mem.vmstat.pgsteal_kswapd_dma32
mem.vmstat.pgsteal_kswapd_movable
mem.vmstat.pgsteal_kswapd_normal
mem.vmstat.pgsteal_movable
mem.vmstat.pgsteal_normal

And this works with RHEL 6 which has:

$ cat /proc/vmstat | grep -E '(pgscan|pgsteal)' | sort
pgscan_direct_dma 0
pgscan_direct_dma32 0
pgscan_direct_movable 0
pgscan_direct_normal 0
pgscan_kswapd_dma 0
pgscan_kswapd_dma32 0
pgscan_kswapd_movable 0
pgscan_kswapd_normal 0
pgsteal_dma 0
pgsteal_dma32 0
pgsteal_movable 0
pgsteal_normal 0

And RHEL 7 which has:

$ cat /proc/vmstat | grep -E '(pgscan|pgsteal)' | sort
pgscan_direct_dma 0
pgscan_direct_dma32 0
pgscan_direct_movable 0
pgscan_direct_normal 0
pgscan_direct_throttle 0
pgscan_kswapd_dma 0
pgscan_kswapd_dma32 0
pgscan_kswapd_movable 0
pgscan_kswapd_normal 0
pgsteal_direct_dma 0
pgsteal_direct_dma32 0
pgsteal_direct_movable 0
pgsteal_direct_normal 0
pgsteal_kswapd_dma 0
pgsteal_kswapd_dma32 0
pgsteal_kswapd_movable 0
pgsteal_kswapd_normal 0

However, on Fedora 27 / kernel 4.13 we have:

$ cat /proc/vmstat | grep -E '(pgscan|pgsteal)' | sort
pgscan_direct 0
pgscan_direct_throttle 0
pgscan_kswapd 0
pgsteal_direct 0
pgsteal_kswapd 0

This means, for example, that stats corresponding to sar -B are impossible to generate with PCP.

Below are few somewhat relevant links for reference:

https://github.com/torvalds/linux/blob/master/include/linux/vm_event_item.h
torvalds/linux@904249a
torvalds/linux@68243e7
torvalds/linux@599d0c9

Please update the PMDA for latest kernel. Thanks.