Blob Blame Raw
From 744e4c6a6cbbb9ba0569bf8e3ab50171e974b2e3 Mon Sep 17 00:00:00 2001
From: Denys Vlasenko <dvlasenk@redhat.com>
Date: Mon, 6 Jan 2014 17:18:31 +0100
Subject: [ABRT PATCH 19/27] Fix handling of Machine Check Exceptions.

Closes #764.

If non-fatal MCE is seen, abrt will detect it as an oops
and alert user in a usual manner. When user opens this
abrt problem for reporting, he will see that "comment"
field is pre-filled with a text.
What it says depends on whether mcelog tool is installed.
If mcelog is installed, the text will say that hardware errors
were detected, and will show the tail of either /var/log/mcelog
or syslog.
Otherwise the text will say that hardware errors
were detected, but they can't be usefully diagnosed,
and user is strongly advised to install mcelog tool.

If fatal MCE is encountered, kernel always panics,
(abrt has no chance of catching the oops),
kdump kicks in, and then after reboot abrt says that new vmcore
is found. When user generates backtrace, he will see oops text
which starts with
"Machine Check Exception: BANK nnn ..." and (hopefully)
is already explanatory enough.

(Yes, it's weird that kernel shows human-readable error messages
on fatal MCEs but doesn't do that for non-fatal ones.
This makes fetching MCE info significantly different...
I wish kernel would show human-readable MCEs in both cases,
we wouldn't need mcelog then... oh well.)

In order to generate meaningful hash for MCE's,
oops hashing was extended for oopses without backtraces.

Since MCEs, unlike regular oopses, don't contain kernel version,
additional magic is added to extract kernel version
in vmcore event handling.

Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>

Related to rhbz#1032077

Signed-off-by: Jakub Filak <jfilak@redhat.com>
---
 src/lib/kernel.c              | 31 +++++++++++++++++++++++++
 src/plugins/koops_event.conf  | 54 +++++++++++++++++++++++++++++++++++++++++++
 src/plugins/vmcore_event.conf | 18 ++++++++++++++-
 3 files changed, 102 insertions(+), 1 deletion(-)

diff --git a/src/lib/kernel.c b/src/lib/kernel.c
index ce8815b..340ec39 100644
--- a/src/lib/kernel.c
+++ b/src/lib/kernel.c
@@ -115,8 +115,29 @@ static const char *const s_koops_suspicious_strings[] = {
      * arch/x86/kernel/cpu/mcheck/p5.c:		"CPU#%d: Machine Check Exception:  0x%8X (type 0x%8X).\n",
      * arch/x86/kernel/cpu/mcheck/mce.c:	pr_emerg(HW_ERR "CPU %d: Machine Check Exception: %Lx Bank %d: %016Lx\n",
      * drivers/edac/sb_edac.c:			printk("CPU %d: Machine Check Exception: %Lx Bank %d: %016Lx\n",
+     *
+     * MCEs can be fatal (they panic kernel) or not.
+     * Fatal MCE are delivered as exception#18 to the CPU.
+     * Non-fatal ones sometimes are delivered as exception#18;
+     * other times they are silently recorded in magic MSRs, CPU is not alerted.
+     * Linux kernel periodically (up to 5 mins interval) reads those MSRs
+     * and if MCE is seen there, it is piped in binary form through
+     * /dev/mcelog to whoever listens on it. (Such as mcelog tool in --daemon
+     * mode; but cat </dev/mcelog would do too).
+     *
+     * "Machine Check Exception:" message is printed *only*
+     * by fatal MCEs (so far, future kernels may be different).
+     * It will be caught as vmcore if kdump is configured.
+     *
+     * Non-fatal MCEs have "[Hardware Error]: Machine check events logged"
+     * message in kernel log.
+     * When /dev/mcelog is read, *no additional kernel log messages appear*:
+     * if we want more readable data, we must rely on other tools
+     * (such as mcelog daemon consuming binary /dev/mcelog and writing
+     * human-readable /var/log/mcelog).
      */
     "Machine Check Exception:",
+    "Machine check events logged",
 
     /* X86 TRAPs */
     "divide error:",
@@ -299,6 +320,16 @@ next_line:
             if (strcasestr(curline, "Call Trace:")) /* yes, it must be case-insensitive */
                 inbacktrace = 1;
             else
+            /* Fatal MCE's have a few lines of useful information between
+             * first "Machine check exception:" line and the final "Kernel panic"
+             * line. Such oops, of course, is only detectable in kdumps (tested)
+             * or possibly pstore-saved logs (I did not try this yet).
+             * In order to capture all these lines, we treat final line
+             * as "backtrace" (which is admittedly a hack):
+             */
+            if (strstr(curline, "Kernel panic - not syncing"))
+                inbacktrace = 1;
+            else
             if (strnlen(curline, 9) > 8
              && (  (curline[0] == '(' && curline[1] == '[' && curline[2] == '<')
                 || (curline[0] == '[' && curline[1] == '<'))
diff --git a/src/plugins/koops_event.conf b/src/plugins/koops_event.conf
index c0277c8..7dfbe36 100644
--- a/src/plugins/koops_event.conf
+++ b/src/plugins/koops_event.conf
@@ -4,6 +4,60 @@ EVENT=post-create analyzer=Kerneloops
         abrt-action-analyze-oops &&
         dmesg >>dmesg &&
         abrt-action-save-kernel-data
+        abrt-action-save-kernel-data || exit $?
+        #
+        # If it exists, we can save a copy of MCE log here:
+        #test -f /var/log/mcelog && cp /var/log/mcelog .
+        # but in current config, sosreport already does that.
+        #
+        # See if MCEs were seen but mcelog isn't installed or running
+        grep -qFi 'Machine check events logged' dmesg || exit 0
+        #
+        # There was an MCE. IOW: it's not a bug, it's a HW error.
+        # Did mcelog logged it to /var/log/mcelog
+        # (RHEL6 by default does this)?
+        test -f /var/log/mcelog &&
+        {
+                # (Ab)use user comment field to inform user about it.
+                echo "The kernel log indicates that hardware errors were detected."
+                echo "/var/log/mcelog file may have more information."
+                echo "The last 20 lines of /var/log/mcelog are:"
+                echo "========================================="
+                # Redirecting sterr in case selinux makes it unreadable
+                # (annoying anyway, but at least user knows what's going on):
+                tail -n20 /var/log/mcelog 2>&1
+                exit 0
+        } >comment
+        #
+        # On RHEL7, mcelog is run so that its output ends up in syslog.
+        # Do we see that?
+        grep -qFi 'mcelog: Hardware event' /var/log/messages &&
+        {
+                echo "The kernel log indicates that hardware errors were detected."
+                echo "System log may have more information."
+                echo "The last 20 mcelog lines of system log are:"
+                echo "========================================="
+                # Redirecting sterr in case selinux makes it unreadable
+                # (annoying anyway, but at least user knows what's going on):
+                grep -Fi 'mcelog:' /var/log/messages | tail -n20 2>&1
+                exit 0
+        } >comment
+        #
+        # Apparently, there is no running mcelog daemon!
+        # Let user know that he needs one.
+        {
+        echo "The kernel log indicates that hardware errors were detected."
+        echo "The data was saved by kernel for processing by the mcelog tool."
+        echo "However, neither /var/log/mcelog nor system log contain mcelog messages."
+        echo "Most likely reason is that mcelog is not installed or not configured"
+        echo "to be started during boot."
+        echo "Without this tool running, the binary data saved by kernel"
+        echo "is of limited usefulness."
+        echo "(You can save this data anyway by running 'cat </dev/mcelog >FILE')."
+        echo "The recommended course of action is to install mcelog."
+        echo "If another hardware error would occur, a user-readable description"
+        echo "of it will be saved in system log or /var/log/mcelog."
+        } >comment
 
 # If you want behavior similar to one provided by kerneloops daemon
 # distributed by kerneloops.org - that is, if you want
diff --git a/src/plugins/vmcore_event.conf b/src/plugins/vmcore_event.conf
index f8de3c5..655d842 100644
--- a/src/plugins/vmcore_event.conf
+++ b/src/plugins/vmcore_event.conf
@@ -1,6 +1,22 @@
 # analyze
 EVENT=analyze_VMcore analyzer=vmcore
-        abrt-action-analyze-vmcore &&
+        # If kdump machinery already extracted dmesg...
+        if test -f vmcore-dmesg.txt; then
+            # ...use that
+            abrt-dump-oops -o vmcore-dmesg.txt >backtrace || exit $?
+            #
+            # Does "kernel" element exist?
+            test -f kernel && exit 0
+            #
+            # Try creating it from vmcore-dmesg.txt:
+            # MCE oopses don't have kernel version in them,
+            # but it should be specified earlier in the log.
+            k=`sed -n '/Linux version/ s/.*Linux version \([^ ]*\) .*/\1/p' vmcore-dmesg.txt | tail -n1`
+            test "$k" != "" && printf "%s" "$k" >kernel
+        else
+            # No vmcore-dmesg.txt, do it the hard way:
+            abrt-action-analyze-vmcore
+        fi &&
         abrt-action-analyze-oops &&
         abrt-action-save-kernel-data
 
-- 
1.8.3.1