Re: [PATCH] RAS/CEC: Add debugfs switch to disable at run time

From: Cong Wang
Date: Sat Apr 20 2019 - 01:43:35 EST


On Thu, Apr 18, 2019 at 5:26 PM Borislav Petkov <bp@xxxxxxxxx> wrote:
>
> Now, if any of that above still doesn't make it clear, please state what
> you're trying to achieve and I'll try to help.

Sorry that I misled you to believe we don't even enable
CONFIG_X86_MCELOG_LEGACY. Here is what we have and
what we have tried:

1. We have CONFIG_X86_MCELOG_LEGACY=y

2. We also have CONFIG_RAS=y and CONFIG_RAS_CEC=y

3. mcelog started as a daemon successfully, like before

4. Some real correctable memory errors happened, as logged in
dmesg

5. mcelog couldn't receive any of them, reported 0 errors

6. Admin's complained to us as they believe this is a kernel bug

7. We dug into kernel source code and found out CONFIG_RAS
hijacks all these errors, by stopping there in the notification chain:

static int mce_first_notifier(struct notifier_block *nb, unsigned long val,
void *data)
{
struct mce *m = (struct mce *)data;

if (!m)
return NOTIFY_DONE;

if (cec_add_mce(m))
return NOTIFY_STOP; // <=== Returns and stops here

/* Emit the trace record: */
trace_mce_record(m);

set_bit(0, &mce_need_notify);

mce_notify_irq(); // <=== There is where MCELOG receives

return NOTIFY_DONE;
}


8. I noticed rasdaemon, and tried to start it instead of mcelog.

9. I injected some memory error and could successfully read them
via ras-mc-ctl.

To demonstrate what I think we should have, here is the PoC code
ONLY to show the idea (please don't judge it):


@ -567,12 +567,12 @@ static int mce_first_notifier(struct
notifier_block *nb, unsigned long val,
void *data)
{
struct mce *m = (struct mce *)data;
+ bool consumed;

if (!m)
return NOTIFY_DONE;

- if (cec_add_mce(m))
- return NOTIFY_STOP;
+ consumed = cec_add_mce(m);

/* Emit the trace record: */
trace_mce_record(m);
@@ -581,7 +581,7 @@ static int mce_first_notifier(struct
notifier_block *nb, unsigned long val,

mce_notify_irq();

- return NOTIFY_DONE;
+ return consumed ? NOTIFY_STOP : NOTIFY_DONE;
}

With this change, although not even compiled, mcelog should still
receive correctable memory errors like before, even when we have
CONFIG_RAS_CEC=y.

Does this make any sense to you?

Thanks!