Re: [PATCH] memcg: add tracing for memcg stat updates

[Shakeel Butt <shakeel.butt@linux.dev>]

On Wed, Oct 09, 2024 at 06:24:55PM GMT, Yosry Ahmed wrote:
> On Wed, Oct 9, 2024 at 6:08 PM Steven Rostedt <rostedt@goodmis.org> wrote:
> >
> > On Wed, 9 Oct 2024 17:46:22 -0700
> > Yosry Ahmed <yosryahmed@google.com> wrote:
> >
> > > > +++ b/mm/memcontrol.c
> > > > @@ -71,6 +71,10 @@
> > > >
> > > >  #include <linux/uaccess.h>
> > > >
> > > > +#define CREATE_TRACE_POINTS
> > > > +#include <trace/events/memcg.h>
> > > > +#undef CREATE_TRACE_POINTS
> > > > +
> > > >  #include <trace/events/vmscan.h>
> > > >
> > > >  struct cgroup_subsys memory_cgrp_subsys __read_mostly;
> > > > @@ -682,7 +686,9 @@ void __mod_memcg_state(struct mem_cgroup *memcg, enum memcg_stat_item idx,
> > > >                 return;
> > > >
> > > >         __this_cpu_add(memcg->vmstats_percpu->state[i], val);
> > > > -       memcg_rstat_updated(memcg, memcg_state_val_in_pages(idx, val));
> > > > +       val = memcg_state_val_in_pages(idx, val);
> > > > +       memcg_rstat_updated(memcg, val);
> > > > +       trace_mod_memcg_state(memcg, idx, val);
> > >
> > > Is it too unreasonable to include the stat name?
> > >
> > > The index has to be correlated with the kernel config and perhaps even
> > > version. It's not a big deal, but if performance is not a concern when
> > > tracing is enabled anyway, maybe we can lookup the name here (or in
> > > TP_fast_assign()).
> >
> > What name? Is it looked up from idx? If so, you can do it on the reading of

Does reading side mean the one reading /sys/kernel/tracing/trace will do
the translation from enums to string?

> > the trace event where performance is not an issue. See the __print_symbolic()
> > and friends in samples/trace_events/trace-events-sample.h
> 
> Yeah they can be found using idx. Thanks for referring us to
> __print_symbolic(), I suppose for this to work we need to construct an
> array of {idx, name}. I think we can replace the existing memory_stats
> and memcg1_stats/memcg1_stat_names arrays with something that we can
> reuse for tracing, so we wouldn't need to consume extra space.
> 
> Shakeel, what do you think?

Cc Daniel & Martin

I was planning to use bpftrace which can use dwarf/btf to convert the
raw int to its enum string. Martin provided the following command to
extract the translation from the kernel.

$ bpftool btf dump file /sys/kernel/btf/vmlinux | grep -A10 node_stat_item
[2264] ENUM 'node_stat_item' encoding=UNSIGNED size=4 vlen=46
        'NR_LRU_BASE' val=0
        'NR_INACTIVE_ANON' val=0
        'NR_ACTIVE_ANON' val=1
        'NR_INACTIVE_FILE' val=2
        'NR_ACTIVE_FILE' val=3
        'NR_UNEVICTABLE' val=4
        'NR_SLAB_RECLAIMABLE_B' val=5
        'NR_SLAB_UNRECLAIMABLE_B' val=6
        'NR_ISOLATED_ANON' val=7
        'NR_ISOLATED_FILE' val=8
...

My point is userspace tools can use existing infra to extract this
information.

However I am not against adding __print_symbolic() (but without any
duplication), so users reading /sys/kernel/tracing/trace directly can
see more useful information as well. Please post a follow up patch after
this one.

thanks for the review,
Shakeel