Re: [PATCH v1] RAS/CEC: Memory Corrected Errors consistent event filtering

From: William Roche
Date: Fri Mar 26 2021 - 18:25:42 EST


On 26/03/2021 20:02, Borislav Petkov wrote:
On Fri, Mar 26, 2021 at 02:30:29PM -0400, “William Roche wrote:
From: William Roche <william.roche@xxxxxxxxxx>

The Corrected Error events collected by the cec_add_elem() have to be
consistently filtered out.
We fix the case where the value of find_elem() to find the slot of a pfn
was mistakenly used as the return value of the function.
Now the MCE notifiers chain relying on MCE_HANDLED_CEC would only report
filtered corrected errors that reached the action threshold.

Signed-off-by: William Roche <william.roche@xxxxxxxxxx>
---

Notes:
Some machines are reporting Corrected Errors events without any
information about a PFN Soft-offlining or Invalid pfn (report given by
the EDAC module or the mcelog daemon).
A research showed that it reflected the first occurrence of a CE error
on the system which should have been filtered by the RAS_CEC component.
We could also notice that if 2 PFNs are impacted by CE errors, the PFN
on the non-zero slot gets its CE errors reported every-time instead of
being filtered out.
This problem has appeared with the introduction of commit
de0e0624d86ff9fc512dedb297f8978698abf21a where the filtering logic has
been modified.
Could you please review this small suggested fix ?
Thanks in advance for any feedback you could have.
William.

drivers/ras/cec.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)

AFAIU, I think you want something like the below untested hunk:

You set it to 0 when it cannot find an element and that "ret = 1" we can
remove because callers don't care about the offlining threshold - the
only caller that looks at its retval wants to know whether it added the
VA successfully to note that it handled the error.

Makes sense?

What we want is to make cec_add_elem() to return !0 value only
when the given pfn triggered an action, so that its callers should
log the error.

This is what I understand from the code comment :

* When an element reaches it's max count of action_threshold, we try to poison
* it by assuming that errors triggered action_threshold times in a single page
* are excessive and that page shouldn't be used anymore. action_threshold is
* initialized to COUNT_MASK which is the maximum.
*
* That error event entry causes cec_add_elem() to return !0 value and thus
* signal to its callers to log the error.

This is not what I understand from your the below suggestion, as find_elem returns the index of the pfn if found.

What I'm expecting from ras_cec is to "hide" CEs until they reach the action threshold where an action is tried against the impacted PFN, and it's now the time to log the error with the entire notifiers chain.

Am I missing something ?


---
diff --git a/drivers/ras/cec.c b/drivers/ras/cec.c
index ddecf25b5dd4..a29994d726d8 100644
--- a/drivers/ras/cec.c
+++ b/drivers/ras/cec.c
@@ -341,6 +341,8 @@ static int cec_add_elem(u64 pfn)
ca->array[to] = pfn << PAGE_SHIFT;
ca->n++;
+
+ ret = 0;
}
/* Add/refresh element generation and increment count */
@@ -363,12 +365,6 @@ static int cec_add_elem(u64 pfn)
del_elem(ca, to);
- /*
- * Return a >0 value to callers, to denote that we've reached
- * the offlining threshold.
- */
- ret = 1;
-
goto unlock;
}
---

Thx.


Thanks again for your feedback.
W.