[RFC,0/1] mm: Track per-task tlb events

Message

Joe Damato Sept. 14, 2022, 1:51 a.m. UTC

Greetings:

TLB shootdown events can be measured on a per-CPU basis by examining
/proc/interrupts. Further information about TLB events can be harvested
from /proc/vmstat if CONFIG_DEBUG_TLBFLUSH is enabled, but this information
is system-wide.

This information is useful, but on a busy system with many tasks it can be
difficult to disambiguate the source of the TLB shootdown events.

Having this information tracked per-task can enable developers to fix or
tweak userland allocators to reduce the number of IPIs and improve
application performance.

This change adds two new fields to task_struct and signal_struct to help
track TLB events:

	- ngtlbflush: number of tlb flushes generated
	- nrtlbflush: number of tlb flushes received

These stats are exported in /proc/[pid]/stat alongside similar metrics
(e.g. min_flt and maj_flt) for analysis.

I've gotten code into kernel networking / drivers before, but I've never
hacked on mm and mm-adjacent code before. Please let me know if there's a
glaring issue and I'll be happy to tweak this code as necessary.

If this seems OK, I'll send an official v1.

Thanks!

Joe Damato (1):
  mm: Add per-task struct tlb counters

 arch/x86/mm/tlb.c            | 2 ++
 fs/proc/array.c              | 9 +++++++++
 include/linux/sched.h        | 6 ++++++
 include/linux/sched/signal.h | 1 +
 kernel/exit.c                | 6 ++++++
 kernel/fork.c                | 1 +
 6 files changed, 25 insertions(+)