Re: Kernel OOPS with partitioned software raid (+ further questions)

From
Neil Brown <hidden>
Date
2006-10-31 01:47:12
Also in
lkml
On Monday October 30, lkml@digadd.de wrote:
Hi all,

I'm running the following software-raid setup:

two raid 0 with two 250GB disks each (sdd1-sdg1) named md_d2 and md_d3
one raid 5 with three 500GB disks (sda2-sdc2) and the two raid0 as
members named md_d5
one raid 1 with 100MB of each of the 500GB disks (sda1-sdc1) named md_d1

The only raid device that actually has a partition table is md_d5. The
other devices are used unpartitioned, which brings me to the first
question: Is it possible to run partitioned and unpartitioned software
raids at the same time?

Back to the topic now after this question. The resulting problem is: due
to the raid5 layout, the partition table of md_d5 is written to where a
partition table on md_d3 would be as well:

[~]>fdisk -l /dev/md_d3

Disk /dev/md_d3: 500.1 GB, 500113211392 bytes
2 heads, 4 sectors/track, 122097952 cylinders
Units = cylinders of 8 * 512 = 4096 bytes

      Device Boot      Start         End      Blocks   Id  System
/dev/md_d3p1               1      244142      976566   83  Linux
/dev/md_d3p2          244143     5126956    19531256   8e  Linux LVM
/dev/md_d3p3         5126957   488279488  1932610128   8e  Linux LVM

Note that the end of md_d3p3 is way beyond the end of the actual device.
Now during boot udev tries to find out about the content of the devices,
using the vol_id program. It checks the various locations for raid
superblocks, lvm superblocks. What happens show the following strace
excerpts:
snip
The kernel above contains a lot of patches (gentoo's hardened sources),
but the same syndrom can be seen with vanilla 2.6.18 or 2.6.19 rc3.

Even if there are likely a dozend workarounds (create a partition table
on the raid 0s one by one and resync; no not rely on raid=part for
autodetection as the raid5 doesn't come up automatically anyway; don't
use vol_id) this should in my oppinion not happen. The points I'd like
to criticize are:
- The partition table read code, which accepts to create the devices
even though they are obviously wrong,
Yeah, I don't like that either.  I tried submitting a patch to fix it
but different people have different ideas about what the 'right'
solution is (do we clip the partition to size, or make it disappear
completely?).
- The partitioned raid device creation code, which creates subdevices
which are larger than the containing device,
This is exactly the same as the above.  The raid code doesn't create
the subdevices.  The partition reading code does that.
- The layer in the kernel that allows the read beyond end of device down
to the raid driver,
Yes... This is generic_make_request in block/ll_rw_blk.c.
It check that the address is within the partition, but it does not
check that the address after mapping is still within the device.
It assumes partitions always fit within devices.
- Most importantly, the raid driver for failing that bad mannered.
Maybe, though I would rather the block layer guaranteed never to send
a request that was out-of-range...
quoted hunk
I honestly didn't look into the other software raid drivers, which are
likely to produce the same result. The attached patch for raid0.c makes
accesses beyond the end of a device into Buffer I/O errors:

xxxxxx Buffer I/O error on device md_d3p3, logical block 483152512

Regards,
Christian
--- raid0.c.orig	2006-10-30 00:12:22.000000000 +0100
+++ raid0.c	2006-10-30 00:14:48.000000000 +0100
@@ -415,6 +415,10 @@
 	chunksize_bits = ffz(~chunk_size);
 	block = bio->bi_sector >> 1;
 	
+	if (block >= mddev->array_size) {
+		bio_endio(bio, bio->bi_size, -EIO);
+		return 0;
+	}
That would work.  Similar fixes needed in linear, raid10, and possible
raid5.

I'll try to push the 'fix it in the block layer' first though.

Thanks,
NeilBeown
Thread (1 message) 1 message, 1 author, last reply 2006-10-31
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help