Re: [PATCH 1/1] alloc_tag: mark inaccurate allocation counters in /proc/allocinfo output

[Suren Baghdasaryan <surenb@google.com>]

On Thu, Sep 11, 2025 at 11:51 AM Yueyang Pan <pyyjason@gmail.com> wrote:
>
> On Thu, Sep 11, 2025 at 11:13:58AM -0700, Suren Baghdasaryan wrote:
> > On Thu, Sep 11, 2025 at 10:35 AM David Wang <00107082@163.com> wrote:
> > >
> > >
> > > At 2025-09-12 01:25:05, "Yueyang Pan" <pyyjason@gmail.com> wrote:
> > > >On Thu, Sep 11, 2025 at 09:18:29AM -0700, Suren Baghdasaryan wrote:
> > > >> On Thu, Sep 11, 2025 at 9:00 AM Usama Arif <usamaarif642@gmail.com> wrote:
> > > >> >
> > > >> >
> > > >> >
> > > >> > On 11/09/2025 16:47, Yueyang Pan wrote:
> > > >> > > On Thu, Sep 11, 2025 at 11:03:50PM +0800, David Wang wrote:
> > > >> > >>
> > > >> > >> At 2025-09-10 07:49:42, "Suren Baghdasaryan" <surenb@google.com> wrote:
> > > >> > >>> While rare, memory allocation profiling can contain inaccurate counters
> > > >> > >>> if slab object extension vector allocation fails. That allocation might
> > > >> > >>> succeed later but prior to that, slab allocations that would have used
> > > >> > >>> that object extension vector will not be accounted for. To indicate
> > > >> > >>> incorrect counters, mark them with an asterisk in the /proc/allocinfo
> > > >> > >>> output.
> > > >> > >>> Bump up /proc/allocinfo version to reflect change in the file format.
> > > >> > >>>
> > > >> > >>> Example output with invalid counters:
> > > >> > >>> allocinfo - version: 2.0
> > > >> > >>>           0        0 arch/x86/kernel/kdebugfs.c:105 func:create_setup_data_nodes
> > > >> > >>>           0        0 arch/x86/kernel/alternative.c:2090 func:alternatives_smp_module_add
> > > >> > >>>          0*       0* arch/x86/kernel/alternative.c:127 func:__its_alloc
> > > >> > >>>           0        0 arch/x86/kernel/fpu/regset.c:160 func:xstateregs_set
> > > >> > >>>           0        0 arch/x86/kernel/fpu/xstate.c:1590 func:fpstate_realloc
> > > >> > >>>           0        0 arch/x86/kernel/cpu/aperfmperf.c:379 func:arch_enable_hybrid_capacity_scale
> > > >> > >>>           0        0 arch/x86/kernel/cpu/amd_cache_disable.c:258 func:init_amd_l3_attrs
> > > >> > >>>      49152*      48* arch/x86/kernel/cpu/mce/core.c:2709 func:mce_device_create
> > > >> > >>>       32768        1 arch/x86/kernel/cpu/mce/genpool.c:132 func:mce_gen_pool_create
> > > >> > >>>           0        0 arch/x86/kernel/cpu/mce/amd.c:1341 func:mce_threshold_create_device
> > > >> > >>>
> > > >> > >>
> > > >> > >> Hi,
> > > >> > >> The changes may  break some client tools, mine included....
> > > >> > >> I don't mind adjusting my tools, but still
> > > >> > >> Is it acceptable  to change
> > > >> > >>       49152*      48* arch/x86/kernel/cpu/mce/core.c:2709 func:mce_device_create
> > > >> > >> to
> > > >> > >>       +49152      +48 arch/x86/kernel/cpu/mce/core.c:2709 func:mce_device_create*
> > > >> > >>
> > > >> > >> The '+' sign make it still standout when view from a terminal, and client tools, not all of them though, might not need any changes.
> > > >> > >> And when client want to filter out inaccurate data items, it could be done by checking the tailing '*" of func name.
> > > >> > >
> > > >> > > I agree with David on this point. We already have monitoring tool built on top
> > > >> > > of this output across meta fleet. Ideally we would like to keep the format of
> > > >> > > of size and calls the same, even for future version, because adding a * will
> > > >> > > change the format from int to str, which leads to change over the regex parser
> > > >> > > many places.
> > > >> > >
> > > >> > > I think simply adding * to the end of function name or filename is sufficient
> > > >> > > as they are already str.
> > > >> > >
> > > >> >
> > > >> > Instead of:
> > > >> >
> > > >> > 49152*      48* arch/x86/kernel/cpu/mce/core.c:2709 func:mce_device_create
> > > >> >
> > > >> > Could we do something like:
> > > >> >
> > > >> > 49152      48 arch/x86/kernel/cpu/mce/core.c:2709 func:mce_device_create(inaccurate)
> > > >>
> > > >> If there is a postprocessing then this would break sometimes later
> > > >> when the function name is parsed, right? So IMO that just postpones
> > > >> the breakage.
> > > >>
> > > >> >
> > > >> > This should hopefully not require any changes to the tools that are consuming this file.
> > > >> > I think it might be better to use "(inaccurate)" (without any space after function name) or
> > > >> > some other text instead of "+" or "*" to prevent breaking such tools. I dont think we need
> > > >> > to even increment allocinfo version number as well then?
> > > >>
> > > >> I'm wondering if we add a new column at the end like this:
> > > >>
> > > >> 49152      48 arch/x86/kernel/cpu/mce/core.c:2709
> > > >> func:mce_device_create [inaccurate]
> > > >>
> > > >> would that break the parsing tools?
> > > >> Well-designed parsers usually throw away additional fields which they
> > > >> don't know how to parse. WDYT?
> > > >>
> > > >
> > > >It would break the parse now as we count the number of string to decide if
> > > >there is an optional module name or not. I don't think it is a big
> > > >deal to fix though.
> >
> > Uh, right. We do have module name as an optional field...
> >
> > >
> > > The inconsistent of module name is really inconvenient for parsing.....
> > > Could we make changes to make it consistent, something like:
> > >
> > > diff --git a/lib/codetag.c b/lib/codetag.c
> > > index 545911cebd25..b8a4595adc95 100644
> > > --- a/lib/codetag.c
> > > +++ b/lib/codetag.c
> > > @@ -124,7 +124,7 @@ void codetag_to_text(struct seq_buf *out, struct codetag *ct)
> > >                                ct->filename, ct->lineno,
> > >                                ct->modname, ct->function);
> > >         else
> > > -               seq_buf_printf(out, "%s:%u func:%s",
> > > +               seq_buf_printf(out, "%s:%u [kernel] func:%s",
> >
> > Yeah, until someone creates a module called "kernel" :)
> > We could keep the name empty like this:
> >
> > +               seq_buf_printf(out, "%s:%u [] func:%s",
> >
> > but I'm not sure that's the best solution.
> >
>
> I guess the best option would be to decide how the format can evolve
> in the future with some rules in comment or doc. But I am sure who
> will decide the rules...
>
> > If we are really concerned about parsers, I could add an ioctl
> > interface to query the counters which are inaccurate. Would that be
> > better?
>
> I think this would be nice as we just need to add functionality on
> top of the parser when we do need it. I guess most of the time we don't
> care about that temporary imprecision.

We don't care about it until it happens :) I think it would be nice to
have a very visible indication that some values are inaccurate. ioclt
would not do that unfortunately...
Another option is to use the "[]" where we currently encode only
module name to express extra info:

No module, accurate counters:
0        0 arch/x86/kernel/kdebugfs.c:105 func:create_setup_data_nodes

Module, accurate counters:
0        0 arch/x86/kernel/kdebugfs.c:105 [my_module]
func:create_setup_data_nodes

No module, inaccurate counters:
0        0 arch/x86/kernel/kdebugfs.c:105 [,inaccurate]
func:create_setup_data_nodes

Module, inaccurate counters:
0        0 arch/x86/kernel/kdebugfs.c:105 [my_module,inaccurate]
func:create_setup_data_nodes

Then the rule for parsers would be that whatever is in [] can be
extended with additional values. If they do not recognize the value,
just ignore it.

>
> >
> > BTW, I have other ideas for ioctls, like filtering-out 0-sized
> > allocations and such.
>
> You mean using ioctl to control if we only print out the non zero
> ones?

Correct.

>
> >
> > >                                ct->filename, ct->lineno, ct->function);
> > >  }
> > >
> > >
> > >
> > >
> > > >
> > > >I think one more important thing is probably to reach a consensus on
> > > >what format can be changed in the future, for example say, we can
> > > >keep adding columns but not change the format the type of one column.
> > > >With such consensus in mind, it will be easier to design the parser.
> > > >And I guess many companies will build parser upon this info for fleet-
> > > >wise collection.
> > > >
> > > >> >
> > > >> > >>
> > > >> > >> (There would be some corner cases, for example, the '+' sign may not needed when the value reach a negative value if some underflow bug happened)
> > > >> > >>
> > > >> > >>
> > > >> > >> Thanks
> > > >> > >> David.
> > > >> > >>
> > > >> > >>
> > > >> > >>> Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
> > > >> > >>> Signed-off-by: Suren Baghdasaryan <surenb@google.com>
> > > >> > >>> ---
> > > >> > >>
> > > >> > >
> > > >> > > Thanks
> > > >> > > Pan
> > > >> >
> > > >
> > > >Thanks
> > > >Pan
>
> Thanks
> Pan