Re: [PATCH v1 07/12] dyndbg: repack struct _ddebug

[Łukasz Bartosik]

niedz., 26 lis 2023 o 07:00 <jim.cromie@xxxxxxxxx> napisał(a):
>
> On Fri, Nov 24, 2023 at 7:39 AM Łukasz Bartosik <lb@xxxxxxxxxxxx> wrote:
> >
> > niedz., 12 lis 2023 o 17:28 Łukasz Bartosik <lb@xxxxxxxxxxxx> napisał(a):
> > >
> > > pt., 10 lis 2023 o 22:01 <jim.cromie@xxxxxxxxx> napisał(a):
> > > >
> > > > On Fri, Nov 10, 2023 at 7:51 AM Łukasz Bartosik <lb@xxxxxxxxxxxx> wrote:
> > > > >
> > > > > sob., 4 lis 2023 o 02:49 <jim.cromie@xxxxxxxxx> napisał(a):
> > > > > >
> > > > > > On Fri, Nov 3, 2023 at 7:10 AM Łukasz Bartosik <lb@xxxxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > From: Jim Cromie <jim.cromie@xxxxxxxxx>
> > > > > > >
> > > > > > > Move the JUMP_LABEL to the top of the struct, since theyre both
> > > > > > > align(8) and this closes a pahole (unfortunately trading for padding,
> > > > > > > but still).
> > > > > > >
> > > > > > > Signed-off-by: Jim Cromie <jim.cromie@xxxxxxxxx>
> > > > > >
> > > > > > let me add, I havent really tested this, nevermind thorough.
> > > > > > specifically, I didnt look for any offset dependence on the static-key
> > > > > > inside their container.
> > > > > > Conversely, maybe theres a free default or something in there.
> > > > > >
> > > > >
> > > > > Any idea how to properly test the relocation of the key ?
> > > >
> > > > I was hoping Jason knew it from memory.
> > > >
> > > > I have booted dd-kitchen-sink, which includes it, and it didnt melt the box.
> > > >
> > > > I just checked `pahole vmlinux` output for the existence of 0-offset keys.
> > > > Its not conclusive, cuz im only looking at x86.
> > > >
> > > > it does occur, but only for "sub-types".
> > > >
> > > > struct static_key_true {
> > > >         struct static_key          key;                  /*     0    16 */
> > > >
> > > >         /* size: 16, cachelines: 1, members: 1 */
> > > >         /* last cacheline: 16 bytes */
> > > > };
> > > > struct static_key_false {
> > > >         struct static_key          key;                  /*     0    16 */
> > > >
> > > >         /* size: 16, cachelines: 1, members: 1 */
> > > >         /* last cacheline: 16 bytes */
> > > > };
> > > > struct static_key_false_deferred {
> > > >         struct static_key_false    key;                  /*     0    16 */
> > > > ...};
> > > > struct static_key_mod {
> > > >         struct static_key_mod *    next;                 /*     0     8 */
> > > > ...};
> > > > struct static_key_deferred {
> > > >         struct static_key          key;                  /*     0    16 */
> > >
> > > I will test it on arm64.
> >
> > Hi Jim,
> >
> > I verified that relocation of JUMP_LABEL to the top of the _ddebug
> > struct does not brak dynamic debug functionality on arm64.
> > I double checked I had CONFIG_JUMP_LABEL enabled in the kernel config for arm64.
> > I was able to enable/disable callsites and see debug logs being written.
> >
> > But if you're concerned there might be issue related to that
> > relocation on other architectures then let's drop this patch
> > and I will use pahole instead of padding for location of flags and
> > trace destination fields.
> > What do you think ?
> >
>
>
> On balance, I think it should go in.
> 0 - my bias was towards abundance of paranoia
> 1 - youve done real work to evaluate the actual risk
> 2 - Jason is on thread, hasnt said WHOA
> 3 - actual patches have seen some testing (lkp-robot included)
> 4 - static-keys/jump-labels have been around a long time
>
> One new topic:
>
> Do you have any thoughts or plans wrt self-testing ?
>

Actually I didn't think about it at all ;). It is not so common
practice to write tests among kernel developers. Addition of trace
instances & events to the dynamic debug is a major change so I see the
value in having it thoroughly tested. That said I'm not saying no to
writing test harness for that purpose but I wonder if there is any
test framework in the kernel that could be reused or is everyone on
their own when it comes to the testing area ?

Thanks,
Lukasz

> the addition of private instances,
> that can be opened & closed, and written to by +T:private_1
>
> would benefit greatly from a test harness to validate it.
> so far all Ive done is demo scripts
>
> :-) thanks
>
> > Thanks,
> > Lukasz