[PATCH] Fix kexec abort due to IPI from panic().
Eric W. Biederman
ebiederm at xmission.com
Thu Sep 16 19:13:38 EDT 2010
Seiji Aguchi <seiji.aguchi at hds.com> writes:
> I'm Seiji Aguchi.
> I work for Hitachi Data Systems.
> It's a first time to send a patch to lkml.
> Nice to meet you.
> I found an issue in kexec.
> Please give me your comments and suggestions.
> Kexec abort when two cpus panic at the same time.
> An example scenario:
> 1. Two cpus panic at the same time .
> 2. One cpu ,cpu0, get kexec_mutex in crash_kexec().
> 3. The other cpu ,cpu1, can't get kexec_mutex and return from crash_kexec().
> 4. Cpu0 runs kmsg_dump(KMSG_DUMP_KEXEC).
> 5. Cpu1 can't get dump_list_lock and return from kmsg_dump(KMSG_DUMP_PANIC).
> 6. Cpu1 runs smp_send_stop() in panic() and sends IPI to other cpus.
> 7. Cpu0 may receive IPI from cpu1 while running kmsg_dump(KMSG_DUMP_KEXEC),
> crash_setup_regs(), or crash_save_vmcore().
> We can solve this issue by disabling external interrupt while getting kexec_mutex
> in crash_kexec().
Disabling interrupts is fine, I thought we did that already at some
point. However that call to kmsg_dump(KMSG_DUMP_KEXEC) is a bug as it
introduces locks into a path that should not be taking locks. Please
remove that broken kmsg_dump call as well.
Nothing in the crash_kexec path should even have the option of blocking.
More information about the kexec