[PATCH] x86, kdump, ioapic: Fix kdump race with migrating irq

Seiji Aguchi seiji.aguchi at hds.com
Mon Feb 20 10:20:34 EST 2012

> -void disable_IO_APIC(void)
> +void disable_IO_APIC(int force)
> {
> 	/*
>+	 * Use force to bust the io_apic spinlock
>+	 *
>+	 * There is a case where kdump can race with irq
>+	 * migration such that kdump will inject an NMI
>+	 * while another cpu holds the ioapic_lock to
>+	 * migrate the irq.  This would cause a deadlock.
>+	 *
>+	 * Because kdump stops all the cpus, we can safely
>+	 * bust the spinlock as we are just clearing the
>+	 * io apic anyway.
>+	 */
>+	if (force && spin_is_locked(&ioapic_lock))
>+		/* only one cpu should be running now */
>+		spin_lock_init(&ioapic_lock);
>+	/*


This patch solves the kdump race, but it may make us confuse when we analyze vmcore,
Because it clears value of ioapic_lock without notice.

It kernel panic is related to IO_APIC, they will check value of ioapic_lock.

So, it is better if we can restore the value of ioapic_lock after calling disable_IO_APIC(),  
IOW, my idea is following.
 - Backup value of ioapic_lock
 - Bust ioapic_lock
 -  Call disable_IO_APIC()
 - Restore value of ioapic_lock

But, as you are discussing with Eric, we don't need to worry about this if we can completely remove disable_IO_APIC().


More information about the kexec mailing list