On 11/21/23 at 09:43am, Ignat Korchagin wrote:
> On Tue, Nov 21, 2023 at 7:53 AM Ignat Korchagin <ignat@xxxxxxxxxxxxxx> wrote:
> >
> > On Tue, Nov 21, 2023 at 1:50 AM Baoquan He <bhe@xxxxxxxxxx> wrote:
> > >
> > > Eric DeVolder's Oracle mail address is not available anymore, add his
> > > current mail address he told me.
> >
> > Thank you!
> >
> > > On 11/20/23 at 10:52pm, Ignat Korchagin wrote:
> > > > Good day!
> > > >
> > > > We have recently started to evaluate Linux 6.6 and noticed that we
> > > > cannot disable CONFIG_KEXEC anymore, but keep CONFIG_CRASH_DUMP
> > > > enabled. It seems to be related to commit 89cde455 ("kexec:
> > > > consolidate kexec and crash options into kernel/Kconfig.kexec"), where
> > > > a CONFIG_KEXEC dependency was added to CONFIG_CRASH_DUMP.
> > > >
> > > > In our current kernel (Linux 6.1) we only enable CONFIG_KEXEC_FILE
> > > > with enforced signature check to support the kernel crash dumping
> > > > functionality and would like to keep CONFIG_KEXEC disabled for
> > > > security reasons [1].
> > > >
> > > > I was reading the long commit message, but the reason for adding
> > > > CONFIG_KEXEC as a dependency for CONFIG_CRASH_DUMP evaded me. And I
> > > > believe from the implementation perspective CONFIG_KEXEC_FILE should
> > > > suffice here (as we successfully used it for crashdumps on Linux 6.1).
> > > >
> > > > Is there a reason for adding this dependency or is it just an
> > > > oversight? Would some solution of requiring either CONFIG_KEXEC or
> > > > CONFIG_KEXEC_FILE work here?
> > >
> > > I searched the patch history, found Eric didn't add the dependency on
> > > CONFIG_KEXEC at the beginning. Later a linux-next building failure with
> > > randconfig was reported, in there CONFIG_CRASH_DUMP enabled, while
> > > CONFIG_KEXEC is disabled. Finally Eric added the KEXEC dependency for
> > > CRASH_DUMP. Please see below link for more details:
> > >
> > > https://lore.kernel.org/all/3e8eecd1-a277-2cfb-690e-5de2eb7b988e@xxxxxxxxxx/T/#u
> >
> > Thank you for digging this up. However I'm still confused, because
> > this is exactly how we configure Linux 6.1 (although we do have
> > CONFIG_KEXEC_FILE enabled) and we don't have any problems. I believe
> > we did not investigate this issue properly.
>
> I did some preliminary investigation for this. If I patch out the
> dependency on CONFIG_KEXEC the kernel builds just fine for x86
> (without CONFIG_CRASH_HOTPLUG - which is probably another issue) - so
> this was the previous behaviour. I can see that the reported error is
> for arm architecture and was able to reproduce it with a simple cross
> compiler in Debian. However, I think it is still somehow related to
> this patchset as the previous kernels (up to 6.5) build fine with just
> CONFIG_CRASH_DUMP and without CONFIG_KEXEC for arm as well. So even
> for arm it was introduced in 6.6.
Thanks for the information.
I haven't run the reproducer of issue reported on Eric's old patchset,
while checkout to kernel 6.1, only s390 selected KEXEC for CRASH_DUMP
already. And with the ARM building breakage, the simplest idea is
to select KEXEC only for ARM or S390 CRASH_DUMP. I plan to try the
reproducer later. If you have any idea or draft patch, please feel free
to post.
diff --git a/kernel/Kconfig.kexec b/kernel/Kconfig.kexec
index 7aff28ded2f4..382dcd8d7a9d 100644
--- a/kernel/Kconfig.kexec
+++ b/kernel/Kconfig.kexec
@@ -97,7 +97,7 @@ config CRASH_DUMP
depends on ARCH_SUPPORTS_KEXEC
select CRASH_CORE
select KEXEC_CORE
- select KEXEC
+ select KEXEC if (ARM || S390)
arch/s390/Kconfig in kernel 6.1:
config CRASH_DUMP
bool "kernel crash dumps"
select KEXEC
help
Generate crash dump after being started by kexec.
Crash dump kernels are loaded in the main kernel with kexec-tools
into a specially reserved region and then later executed after
a crash by kdump/kexec.
Refer to <file:Documentation/s390/zfcpdump.rst> for more details on this.
This option also enables s390 zfcpdump.
See also <file:Documentation/s390/zfcpdump.rst>
>
> > > And besides, the newly added CONFIG_CRASH_HOTPLUG also needs
> > > CONFIG_KEXEC if the elfcorehdr is allowed to be manipulated when
> > > cpu/memory hotplug hapened.
> >
> > This still feels like a regression to me: any crash dump support
> > should be independent of KEXEC syscalls being present. While probably
> > the common case (including us) that the crashing kernel and recovery
> > kernel are the same, they don't have to be. We need kexec syscall in
> > the crashing kernel, but crashdump support in the recovery kernel (but
> > the recovery kernel not having the kexec syscalls should be totally
> > fine). If we do require some code definitions from kexec - at most we
> > should put them under CONFIG_KEXEC_CORE.
> >
> > > Thanks
> > > Baoquan
> > >
>
> Ignat
>
