Hi Tyler,
On 05/01/17 22:31, Baicar, Tyler wrote:
> On 12/20/2016 8:29 AM, James Morse wrote:
>> On 07/12/16 21:48, Tyler Baicar wrote:
>>> ARM APEI extension proposal added SEA (Synchrounous External
>>> Abort) notification type for ARMv8.
>>> Add a new GHES error source handling function for SEA. If an error
>>> source's notification type is SEA, then this function can be registered
>>> into the SEA exception handler. That way GHES will parse and report
>>> SEA exceptions when they occur.
>>> diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c
>>> index 2acbc60..66ab3fd 100644
>>> --- a/drivers/acpi/apei/ghes.c
>>> +++ b/drivers/acpi/apei/ghes.c
>>> @@ -767,6 +771,62 @@ static struct notifier_block ghes_notifier_sci = {
>>> .notifier_call = ghes_notify_sci,
>>> };
>>> +#ifdef CONFIG_HAVE_ACPI_APEI_SEA
>>> +static LIST_HEAD(ghes_sea);
>>> +
>>> +static int ghes_notify_sea(struct notifier_block *this,
>>> + unsigned long event, void *data)
>>> +{
>>> + struct ghes *ghes;
>>> + int ret = NOTIFY_DONE;
>>> +
>>> + rcu_read_lock();
>>> + list_for_each_entry_rcu(ghes, &ghes_sea, list) {
>>> + if (!ghes_proc(ghes))
>>> + ret = NOTIFY_OK;
>>> + }
>>> + rcu_read_unlock();
>>> +
>>> + return ret;
>>> +}
>> What stops this from being re-entrant?
>>
>> ghes_copy_tofrom_phs() takes the ghes_ioremap_lock_irq spinlock, but there is
>> nothing to stop a subsequent instruction fetch or memory access causing another
>> (maybe different) Synchronous External Abort which deadlocks trying to take the
>> same lock.
>>
>> ghes_notify_sea() looks to be based on ghes_notify_sci(), which (if I've found
>> the right part of the ACPI spec) is a level-low interrupt. spin_lock_irqsave()
>> would mask interrupts so there is no risk of a different notification firing on
>> the same CPU, (it looks like they are almost all ultimately an irq).
>>
>> NMI is the odd one out because its not maskable like this, but ghes_notify_nmi()
>> has:
>>> if (!atomic_add_unless(&ghes_in_nmi, 1, 1))
>>> return ret;
>> To ensure there is only ever one thread poking around in this code.
>>
>> What happens if a system describes two GHES sources, one using an irq the other
>> SEA? The SEA error can interrupt the irq error while its holding the above lock.
>> I guess this is also why all the NMI code in that file is separate.
> Let me see if I'm following you right :)
> I should use spin_lock_irqsave() in ghes_notify_sea() to avoid ghes_notify_sci()
> from
> interrupting this process and potentially causing the deadlock?
This way round you are already safe: The CPU masks interrupts when it takes the
exception, they should still be masked by the time we get in here...
The other way round is a lot more fun!
What happens if APEI is processing some error record that was notified via an
interrupt, and then takes the Synchronous External Abort, and ends up back in
this code? Masking interrupts doesn't stop the external-abort, and trying to
take the ghes_ioremap_lock_irq will deadlock.
What happens if we interrupt printk() holding all its locks is another thing I
haven't worked out yet.
> This race condition does seem valid. We are using the same acknowledgment for
> all our
> HEST table entries, so our firmware will not populate more than one entry at a
> time. That
> gets us around this race condition.
Ah, so your firmware will wait for the interrupt-signalled error to be finished
before it triggers the Synchronous External Abort. I think this would still be a
linux bug if the firmware didn't do this.
x86 could have done the same with NMI notifications, but we have all this 'if
(in_nmi)' to allow interrupts-masked GHES handling to be interrupted.
What do you think to re-using the 'if (in_nmi)' code for SEA? We can argue that
SEA is NMI-like in that it can't be masked, and it interrupts code that had
interrupts masked. It 'should' be as simple as putting 'HAVE_NMI' in arm64's
Kconfig, and wrapping the atomic notifier call with nmi_enter()/nmi_exit() from
linux/hardirq.h. (...famous last words...)
This probably answers my printk() questions too, but I need to look into it some
more.
Thanks,
James
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
