Bug 7842 - in-kernel stex driver locks up system, problem with common block layer code?
Summary: in-kernel stex driver locks up system, problem with common block layer code?
Status: CLOSED OBSOLETE
Alias: None
Product: IO/Storage
Classification: Unclassified
Component: SCSI (show other bugs)
Hardware: i386 Linux
: P2 normal
Assignee: Ed Lin
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2007-01-17 08:18 UTC by Thomas Orgis
Modified: 2012-05-17 14:21 UTC (History)
4 users (show)

See Also:
Kernel Version: 2.6.19.1
Subsystem:
Regression: No
Bisected commit-id:


Attachments
trace of the last seen kernel BUG (2.79 KB, text/plain)
2007-01-17 08:20 UTC, Thomas Orgis
Details
kernel messages from syslog with a crash/freeze (24.46 KB, text/plain)
2007-01-17 08:22 UTC, Thomas Orgis
Details
Fix bad sharing of tag busy list (2.20 KB, patch)
2007-10-25 01:14 UTC, Jens Axboe
Details | Diff
2.6.23.1 variant of the patch (2.31 KB, patch)
2007-10-25 01:21 UTC, Jens Axboe
Details | Diff

Description Thomas Orgis 2007-01-17 08:18:55 UTC
Most recent kernel where this bug did *NOT* occur: supposedly any kernel with 
the external Promise stex driver (<2.6.19), please see details

Distribution: OpenSuSE 10.2 (x86_64 smp)
Hardware Environment: Tyan S3970 w/1 Opteron 2210 dual core, 2GB DDR2@667MHz, 
Promise SuperTrak EX 16350/16300 SATA RAID controller w/ 14 disks attached (10 
in RAID10, 4 in RAID5); OS installed on extra HDD connected to mainboard

Software Environment: stock SuSE with vanilla kernel 2.6.19.1, setup as NFS 
server
Problem Description:

I am filing this under SCSI, as the stex driver resides there, though it is a 
SATA device and the actual problem could even be in more generic code.

- installed 2.6.19.1 since stock kernel does not have stex driver
- System or just the filesystems on the RAIDs freeze after some time / amount 
of data processed (normally total freeze, one time the box was still 
responsive, but any access to the RAID mounts was blocking)
- tried different configuration of kernel (no -Os, p.ex.), added debugging... 
effect stays the same

I was able to get two Kernel BUG messages (I'll attach the detailed messages):

Kernel BUG at block/elevator.c:332
invalid opcode: 0000 [1] SMP
CPU 0

Kernel BUG at block/ll_rw_blk.c:1130
invalid opcode: 0000 [1] SMP
CPU 0

I contacted Promise and they provided an updated version of their external 
stex/shasta driver source that compiles against the 2.6.18.2-like kernel of 
OpenSuSE 10.2. This driver works under varying load (also stressed it) for 9 
days now.
There seems to be a problem with the changes the driver did undergo to be 
included in the kernel (replacing code from Promise with using in-kernel common 
code) or with the used common kernel code itself...

That this setup now works reliably (I hope;-) is an indication that switching 
to the external driver is the solution and it's indeed a software problem.
However, I won't hide that I see a correlation between RAID usage and the 
occurence of MCEs regarding corrected ECC errors. Not with CPU/Memory load 
alone -- there seems to be the I/O / PCI-X / whatever action needed.

Steps to reproduce:

- get Promise STEX163x0 card and perhaps an SMP box ...
- configure two RAID sets
- copy big files from one to the other -- that reliably triggered a system 
freeze before reaching 100G of copied data for me
Comment 1 Thomas Orgis 2007-01-17 08:20:39 UTC
Created attachment 10105 [details]
trace of the last seen kernel BUG
Comment 2 Thomas Orgis 2007-01-17 08:22:33 UTC
Created attachment 10106 [details]
kernel messages from syslog with a crash/freeze
Comment 3 Christian Sch 2007-05-22 05:52:20 UTC
Same problem here. The system freezes sporadically.

System:
AMD X2 Dualcore 2GHz, 4GB DDR RAM
Tyan Mainboard S2865 Tomcat K8E
Promise EX 8350 Raid
5 Western Digital HDs (160GB) in one Raid5 (OS + Data)

I'm using Ubuntu 7.04 with Kernel 2.6.20.

The Server ist used as VMware-Server.

Same error messages as above.

Comment 4 Andrew Morton 2007-07-24 12:25:37 UTC
Christian, are you also using the driver from Promise's site
or are you referring to the kernel.org driver?

If it's the external driver then I'll close the bug, sorry - there
isn't anything we can do about it.
Comment 5 Thomas Orgis 2007-07-25 02:31:58 UTC
Please note that it is the kernel.org driver (be it in OpenSUSE or Ubuntu variation) that is causing the freezes for Christian and me; we both got around the problem by using the external driver.

Short:
kernel.org -> freeze
external -> works

Just in case that info got mixed up... 
Also, Christian informed me about a patch the driver guy from Promise (Ed Lin) posted to LKM:

http://lkml.org/lkml/2007/1/23/268

There were two comments on that post and no further discussion.
Perhaps one could revive this one... and generally: Ed, are you with us on this bugzilla entry?
Comment 6 Ed Lin 2007-07-25 11:49:57 UTC
The patch should be still valid technically. Maybe need a little re-work. However, I must be careful to make this patch be accepted. I will re-submit it again in the near future, and hopefully it can get into some 2.6.23-rcx.
Comment 7 Jens Axboe 2007-09-14 02:25:14 UTC
See

http://git.kernel.dk/?p=linux-2.6.git;a=commit;h=f3da54ba140c6427fa4a32913e1bf406f41b5dda

which is now in Linus tree, it should fix this issue.
Comment 8 Jens Axboe 2007-09-14 02:28:25 UTC
This issue may be closed.
Comment 9 Lorenzo 2007-10-17 07:37:07 UTC
Hi.

I use stex driver with kernel 2.6.21-r4 (gentoo).

Kernel hangs when I make heavy i/o operations with scsi volume.
This is the error:

Process kblockd/1 (pid:46, ti=c19e6000 task=c19be530 task.ti=c19e6000)
STACK: C1AE6448 D752F578 F7E5BBA4 C1A7E800 C1A7880 F7E5BBA4 C01CC54B
C1A78800
00000287 C027A1CC C0DDA080 00000000 C1A7E800 C1A78800 C027E98E C1979764
F7E5BBA4 C1979740 C1A7E884 D752F578 FE5BBA4 C1979740 F7E5BC30 00000292

CALL TRACE:
[<C01CC54B>] ELV_NEXT_REQUEST+0x18/0x125
C027A1CC SCSI_DISPATCH_CMD+0x13c/0X215
C027E98E SCSI_REQUEST_FN+0x186/0x269
C01CF273 __GENERIC_UNPLUG_DEVICE+0x21/0x23
C01CFFBE GENERIC_UNPLUG_DEVICE+0x15/0x21
C01CD177 BLK_UNPLUG_WORK+0xB/0xC
C0127A9E RUN_WORK_QUEUE+0x94/0x13F
C01CD16C BLK_UNPLUG_WORK+0x0/0xC
C0128188 WORKER_THREAD+0x143/0x168
C01150B1 DEFAULT_WAKE_FUNCTION+0x0/0xC
C0128045 WORKER_THREAD+0x0/0x168
C012A8F2 KTHREAD+0xAE/0xD3
C012A844 KTHREAD+0x0/0xD3
C01032DB KERNEL_THREAD_HELPER+0x7/0x1C
CODE: C9 74 28 85 D2 74 10 8B 43 2C 8B 4B 3C 8D 50
01 89 53 2C 39 C8 73 9F 31 C9 EB AE C7 43 20
00 00 00 00 8B 54 CB 08 E9 56 FF FF FF <0F>
0B EB FE 0F 0B EB FE 8B 40 0C 8B 50 04 83
C2 0B 8D 42 08 31

EIP:[<C01D3576>] DEADLINE_DISPATCH_REQUEST+0xF3/0xFB



Do you think that the 2.6.23 kernel (and the Ed's patch included) resolve this bug?
Is these errors caused by stex driver?
Thanks.
Comment 10 Andrew Morton 2007-10-17 13:02:58 UTC
A lot can change between 2.6.21 and 2.6.23.  Yes, please do test 2.6.23.
Comment 11 Jens Axboe 2007-10-17 13:10:00 UTC
It's almost 100% certainly the same bug, so yes is fixed.
Comment 12 Lorenzo 2007-10-23 01:05:17 UTC
Updated with 2.6.23 kernel and in-kernel stex driver, same error. (see post #9).
Nothing is changed with new kernel.

I don't know why this driver isn't signed as "experimental".
Comment 13 Lorenzo 2007-10-23 01:15:41 UTC
Updated with 2.6.23 kernel and in-kernel stex driver, same error. (see post #9).
Nothing is changed with new kernel.

I don't know why this driver isn't signed as "experimental".
Comment 14 Thomas Orgis 2007-10-23 01:57:26 UTC
I am wondering if Lorenzo really is talking about the same issue I had.
Sadly, I cannot test the current in-kernel driver since we only have one stex system which is needed in production and there is no further backup space for the many 100s of GB on the RAID:-(

Our box is running fine with the external driver from Promise.
Comment 15 Lorenzo 2007-10-23 02:40:56 UTC
Hi Thomas.

I read that your box is running fine with the external driver from Promise.
Which kernel version is running? Which Promise driver have you selected?
Can you post your .config?

Thank you
Comment 16 Thomas Orgis 2007-10-23 03:10:26 UTC
atlas:~ # uname -a
Linux atlas 2.6.18.2-34-thor #5 SMP Sun Jan 7 15:13:46 CET 2007 x86_64 x86_64 x86_64 GNU/Linux

We have the 2.6.18.2 Kernel from OpenSuse 10.2; I took the OpenSuse kernel sources with the default (smp) config... modified a bit...

The important part is that Ed Lin sent me an updated version of the Promise driver that I then compiled externally.
I recommend you ask him or check out the promise site:

http://www.promise.com/upload/Support/Driver/SuperTrak-EX-Series-suse-10.2-x86_64-2.9.0.22.tgz

This looks like the version I got.
Comment 17 Jens Axboe 2007-10-25 01:14:23 UTC
Created attachment 13271 [details]
Fix bad sharing of tag busy list

I think there's one more bug there, for shared maps. For the locking to
work, only the tag map and tag bit map may be shared (incidentally, I
was just explaining this to Nick yesterday, but I apparently didn't
review the code well enough myself). But we also share the busy list!
The busy_list must be queue private, or we need a block_queue_tag
covering lock as well.

So we have to move the busy_list to the queue. This'll work fine, and
it'll actually also fix a problem with blk_queue_invalidate_tags() which
will invalidate tags across all shared queues. This is a bit confusing,
the low level driver should call it for each queue seperately since
otherwise you cannot kill tags on just a single queue for eg a hard
drive that stops responding. Since the function has no callers
currently, it's not an issue.

Please test this patch, it should fix the issue!
Comment 18 Jens Axboe 2007-10-25 01:21:22 UTC
Created attachment 13272 [details]
2.6.23.1 variant of the patch

Same as 13271, just applies cleanly for 2.6.23.1 for testing there. Lorenzo, please test?
Comment 19 Jens Axboe 2007-11-05 01:27:34 UTC
Ed verified the patch, we should close this bug report. Fix is in 2.6.24-git and is also queued for 2.6.23.x stable series.
Comment 20 Christian Sch 2008-03-10 02:00:22 UTC
I guess it's not solved yet.
I gave it a try and tested kernel 2.6.22.14 (which is the Ubuntu 7.10 kernel) and compiled 2.6.24.3 myself. Both crashed when I made heavy IO. In fact it needed only a ftp transfer of a single 700MB file (GBit ethernet) to crash.
Comment 21 Christian Sch 2008-03-11 05:02:35 UTC
One thing I forgot to tell: I used the 64 bit version. I didn't test any 32 bit kernel.
Comment 22 Alan 2012-05-17 14:21:26 UTC
Closing obsolete bugs, if still relevant please re-open

Note You need to log in before you can comment on or make changes to this bug.