thr3ads.net - Lustre devel - [Patch] Fix Client Kernel Crash with Mis-configured Index Numbering [Oct 2013]

If this information is useful, please help other people find it:
Share via:

Wendy Cheng

2013-Oct-12 03:28 UTC

[Patch] Fix Client Kernel Crash with Mis-configured Index Numbering

Ref: http://lists.lustre.org/pipermail/lustre-devel/2013-October/004270.html

I''m not really convinced the "index" setting of mkfs.lustre
needs to
be started with "0". However, in the minimum, the client kernel should
not crash. The attached patch does this minimum fix; compiled and
tested with GIT master branch.

Recreated by:
server> mkfs.lustre --reformat --fsname=lus1 --mgs --mdt --index=1 /dev/sdd1
server> mkfs.lustre --reformat --ost --fsname=lus1
--mgsnode=192.168.20.46@o2ib0 --index=1 /dev/sde1

client> mount.lustre -o flock 192.168.20.46@o2ib0:/lus1 /mnt/lustre

The client mount crashes at lmv_get_info() without changes

<1>[  215.946538] BUG: unable to handle kernel NULL pointer
dereference at 0000000000000028
<1>[  215.946572] IP: [<ffffffffa07445cb>] lmv_get_info+0x32b/0x560
[lmv]
<0>[  215.947090] Call Trace:^M
<4>[  215.947143]  [<ffffffffa0655b70>] ll_fill_super+0x1f40/0x4330
[lustre]^M
<4>[  215.947214]  [<ffffffffa02cf527>] ?
lustre_start_mgc+0x227/0x2a90 [obdclass]^M
<4>[  215.947275]  [<ffffffffa02d3d60>]
lustre_fill_super+0xa20/0x22f0
[obdclass]^M
<4>[  215.947304]  [<ffffffff810de91f>] ? set_anon_super+0x0/0xe0^M
<4>[  215.947361]  [<ffffffffa02d3340>] ?
lustre_fill_super+0x0/0x22f0
[obdclass]^M
<4>[  215.947380]  [<ffffffff810df601>] mount_nodev+0x50/0x84^M
<4>[  215.947437]  [<ffffffffa02cc5d9>] lustre_mount+0x29/0x30
[obdclass]^M
<4>[  215.947454]  [<ffffffff810df009>] vfs_kern_mount+0xa8/0x1f3^M
<4>[  215.947471]  [<ffffffff810df1bc>] do_kern_mount+0x4d/0xe1^M
<4>[  215.947489]  [<ffffffff810f54d7>] do_mount+0x67d/0x6d5^M
<4>[  215.947507]  [<ffffffff810f57cc>] sys_mount+0x84/0xbd^M
<4>[  215.947527]  [<ffffffff81002aab>]
system_call_fastpath+0x16/0x1b^M

Signed-off-by: Wendy Cheng
<wendy.cheng-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>

diff --git a/lustre/lmv/lmv_obd.c b/lustre/lmv/lmv_obd.c
index 3091bfb..5f4a18b 100644
--- a/lustre/lmv/lmv_obd.c
+++ b/lustre/lmv/lmv_obd.c
@@ -2443,6 +2443,16 @@ static int lmv_get_info(const struct lu_env
*env, struct obd_export *exp,
                         RETURN(rc);

                /*
+                * In the case of mis-configured OSS, instead of crashing
+                * the kernel during client mount, give them a warning and
+                * gracefully back out mount process w/ -ENXIO error.
+                */
+               if (lmv->tgts[0] == NULL) {
+                       CDEBUG(D_IOCTL, "NULL index\n");
+                       RETURN(-ENXIO);
+               }
+
+               /*
                 * Forwarding this request to first MDS, it should know LOV
                 * desc.
                 */

Dilger, Andreas

2013-Oct-12 06:59 UTC

head link

Re: [Patch] Fix Client Kernel Crash with Mis-configured Index Numbering

Hi Wendy,
Thanks for the patch. Could you please file a ticket at
https://jira.hpdd.intel.com/ and submit the patch to our Gerrit repo (with minor
tweaks as suggested below) so it is included in the next Lustre release. For
more details please see:

https://wiki.hpdd.intel.com/display/PUB/Submitting+Changes

You are totally correct that no user input should crash the kernel. The support
for multiple MDTs in the same filesystem is relatively new (previously only MDT
index 0 was allowed), and I guess nobody has ever tested what you did.

Cheers, Andreas

On 2013-10-11, at 21:29, "Wendy Cheng"
<s.wendy.cheng-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org<mailto:s.wendy.cheng-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>>
wrote:

Ref: http://lists.lustre.org/pipermail/lustre-devel/2013-October/004270.html

I''m not really convinced the "index" setting of mkfs.lustre
needs to
be started with "0". However, in the minimum, the client kernel should
not crash. The attached patch does this minimum fix; compiled and
tested with GIT master branch.

Recreated by:
server> mkfs.lustre --reformat --fsname=lus1 --mgs --mdt --index=1 /dev/sdd1
server> mkfs.lustre --reformat --ost --fsname=lus1
--mgsnode=192.168.20.46@o2ib0 --index=1 /dev/sde1

client> mount.lustre -o flock 192.168.20.46@o2ib0:/lus1 /mnt/lustre

The client mount crashes at lmv_get_info() without changes

<1>[  215.946538] BUG: unable to handle kernel NULL pointer
dereference at 0000000000000028
<1>[  215.946572] IP: [<ffffffffa07445cb>] lmv_get_info+0x32b/0x560
[lmv]
<0>[  215.947090] Call Trace:^M
<4>[  215.947143]  [<ffffffffa0655b70>] ll_fill_super+0x1f40/0x4330
[lustre]^M
<4>[  215.947214]  [<ffffffffa02cf527>] ?
lustre_start_mgc+0x227/0x2a90 [obdclass]^M
<4>[  215.947275]  [<ffffffffa02d3d60>]
lustre_fill_super+0xa20/0x22f0
[obdclass]^M
<4>[  215.947304]  [<ffffffff810de91f>] ? set_anon_super+0x0/0xe0^M
<4>[  215.947361]  [<ffffffffa02d3340>] ?
lustre_fill_super+0x0/0x22f0
[obdclass]^M
<4>[  215.947380]  [<ffffffff810df601>] mount_nodev+0x50/0x84^M
<4>[  215.947437]  [<ffffffffa02cc5d9>] lustre_mount+0x29/0x30
[obdclass]^M
<4>[  215.947454]  [<ffffffff810df009>] vfs_kern_mount+0xa8/0x1f3^M
<4>[  215.947471]  [<ffffffff810df1bc>] do_kern_mount+0x4d/0xe1^M
<4>[  215.947489]  [<ffffffff810f54d7>] do_mount+0x67d/0x6d5^M
<4>[  215.947507]  [<ffffffff810f57cc>] sys_mount+0x84/0xbd^M
<4>[  215.947527]  [<ffffffff81002aab>]
system_call_fastpath+0x16/0x1b^M

Signed-off-by: Wendy Cheng
<wendy.cheng-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org<mailto:wendy.cheng-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>>

diff --git a/lustre/lmv/lmv_obd.c b/lustre/lmv/lmv_obd.c
index 3091bfb..5f4a18b 100644
--- a/lustre/lmv/lmv_obd.c
+++ b/lustre/lmv/lmv_obd.c
@@ -2443,6 +2443,16 @@ static int lmv_get_info(const struct lu_env
*env, struct obd_export *exp,
                        RETURN(rc);

               /*
+                * In the case of mis-configured OSS, instead of crashing

This comment should read "misconfigured MDT" ...

+                * the kernel during client mount, give them a warning and
+                * gracefully back out mount process w/ -ENXIO error.
+                */
+               if (lmv->tgts[0] == NULL) {
+                       CDEBUG(D_IOCTL, "NULL index\n");

"NULL target for MDT0\n"

+                       RETURN(-ENXIO);
+               }
+
+               /*
                * Forwarding this request to first MDS, it should know LOV
                * desc.
                */
_______________________________________________
Lustre-devel mailing list
Lustre-devel-aLEFhgZF4x6X6Mz3xDxJMA@public.gmane.org<mailto:Lustre-devel-aLEFhgZF4x6X6Mz3xDxJMA@public.gmane.org>
http://lists.lustre.org/mailman/listinfo/lustre-devel

Wendy Cheng

2013-Oct-13 13:42 UTC

head link

Re: [Patch] Fix Client Kernel Crash with Mis-configured Index Numbering

On Fri, Oct 11, 2013 at 11:59 PM, Dilger, Andreas
<andreas.dilger-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
wrote:> Hi Wendy,
> Thanks for the patch. Could you please file a ticket at
> https://jira.hpdd.intel.com/ and submit the patch to our Gerrit repo (with
> minor tweaks as suggested below) so it is included in the next Lustre
> release. For more details please see:
>
> https://wiki.hpdd.intel.com/display/PUB/Submitting+Changes
>
>
Thanks ... Will do the check in sometime next week when I''m back to
office .

-- Wendy

Lustre devel - Oct 2013 - [Patch] Fix Client Kernel Crash with Mis-configured Index Numbering

[Patch] Fix Client Kernel Crash with Mis-configured Index Numbering

Re: [Patch] Fix Client Kernel Crash with Mis-configured Index Numbering

Re: [Patch] Fix Client Kernel Crash with Mis-configured Index Numbering