On Tue 2026-01-13 21:32:33, Marcos Paulo de Souza wrote:
> On Tue, 2026-01-13 at 09:41 -0300, Marcos Paulo de Souza wrote:
> > On Mon, 2026-01-05 at 14:08 +0000, Daniel Thompson wrote:
> > > On Mon, Jan 05, 2026 at 12:52:14PM +0000, Daniel Thompson wrote:
> > > > Hi Marcos
> > > > 
> > > > On Sat, Dec 27, 2025 at 09:16:07AM -0300, Marcos Paulo de Souza
> > > > wrote:
> > > > > The parts 1 and 2 can be found here [1] and here[2].
> > > > > 
> > > > > The changes proposed in this part 3 are mostly to clarify the
> > > > > usage of
> > > > > the interfaces for NBCON, and use the printk helpers more
> > > > > broadly.
> > > > > Besides it, it also introduces a new way to register consoles
> > > > > and drop thes the CON_ENABLED flag. It seems too much, but in
> > > > > reality
> > > > > the changes are not complex, and as the title says, it's
> > > > > basically a
> > > > > cleanup without changing the functional changes.
> > > > 
> > > > I ran this patchset through the kgdb test suite and I'm afraid it
> > > > is
> > > > reporting functional changes.
> > > > 
> > > > Specifically the earlycon support for kdb has regressed (FWIW the
> > > > problem bisects down to the final patch in the series where
> > > > CON_ENABLED
> > > > is removed).
> > > > 
> > > > Reproduction on x86-64 KVM outside of the test suite should be
> > > > easy:
> > > > 
> > > >     make defconfig
> > > >     scripts/config \
> > > >         --enable DEBUG_INFO \
> > > >         --enable DEBUG_INFO_DWARF_TOOLCHAIN_DEFAULT \
> > > >         --enable DEBUG_FS \
> > > >         --enable KALLSYMS_ALL \
> > > >         --enable MAGIC_SYSRQ \
> > > >         --enable KGDB \
> > > >         --enable KGDB_TESTS \
> > > >         --enable KGDB_KDB \
> > > >         --enable KDB_KEYBOARD \
> > > >         --enable LKDTM \
> > > >         --enable SECURITY_LOCKDOWN_LSM
> > > >     make olddefconfig
> > > >     make -j$(nproc)
> > > >     qemu-system-x86_64 \
> > > >         -m 1G -smp 2 -nographic \
> > > >         -kernel arch/x86/boot/bzImage \
> > > >         -append "console=ttyS0,115200 kgdboc=ttyS0
> > > > earlycon=uart8250,io,0x3f8 kgdboc_earlycon kgdbwait"
> > > 
> > > Actually I realized there was a simpler reproduction (hinted at by
> > > the
> > > missing "printk: legacy bootconsole [uart8250] enabled" in the
> > > regressed
> > > case). It looks like the earlycon simply doesn't work and that
> > > means
> > > the
> > > reproduction doesn't require anything related to kgdb at all.
> > > Simply:
> > > 
> > >     make defconfig
> > >     make -j$(nproc)
> > >     qemu-system-x86_64 -m 1G -smp 2 -nographic -kernel
> > > arch/x86/boot/bzImage \
> > >         -append "earlycon=uart8250,io,0x3f8"
> > > 
> > > With the part 3 patchset applied I get no output from the earlycon
> > > (without the patch set I get the early boot messages which, as
> > > expected,
> > > stop when tty0 comes up).
> > 
> > Hi Daniel, sorry for the late reply! Lots of things to check lately
> > :)
> > 
> > Ok, I reproduced here, thanks a lot for testing kgdboc, it's a quick
> > way to check that the new register_console_force is not working. Let
> > me
> > take a look to find what's wrong. Thanks a lot for finding this
> > issue!
> 
> Ok, I did a bisect and found out that the issue lies in the last
> commit, where CON_ENABLED was removed. After it, I then checked what
> was wrong, since everything was being plumbed correctly (tm), and then
> I found that it was not:
> 
> On _register_console, the function try_enable_default_console is called
> when there are not registered consoles, and then it sets CON_ENABLED
> for the console. Later on, try_enable_preferred_console it checks if
> the console was specified by the user, and at the same time it had
> CON_ENABLED set.
> 
> It worked by chance, but now, we don't have this flag anymore, and then
> we are not _marking_ the console on try_enable_default_console so
> try_enable_preferred_console returns ENOENT.

Great catch! Yeah, it worked just by chance.

> So, without any console kgdb is activated much later in the boot
> process, as you found it.
> 
> I talked with Petr Mladek and it would need to rework the way that we
> register a console, and he's already working on it.

Yes, I have some patches in early stages of developnent of another
feature which would help here.

> For now I believe
> that we could take a look in all the patches besides the last one that
> currently breaks the earlycon with kgdb and maybe other usecases.

I agree. I am going to review this patchset first. Then I'll try to
clean up the patches which remove the ugly side effect from
try_enable_preferred_console(). Then we could discuss how
to move forward. It might make sense to push this patchset
first without the last patch...

> Sorry for not catching this issue before. I'll use kgdb next time to
> make sure that it keeps working :)

Do not worry at all. It was a well hidden catch. It is great that
Daniel found the regression in time...

Best Regards,
Petr

Reply via email to