RedHat 9 aacraid - system fails under extreme disk IO - Repro ducable test case

Andrew Mann amann at mythicentertainment.com
Wed Oct 8 12:39:00 CDT 2003


Hi Mark,
	I've got some potentially interesting results on this.
I downloaded kernel-source-2.4.20.SuSE-62.src.rpm from the Suse ftp 
site.  It's laid out very nice for this purpose.  It's separated into 
the stock 2.4.20 kernel with patches for each arch (and a common patch 
set).  Inside of the common patch set are patches to the aacraid driver. 
  I applied these patches to the stock 2.4.20 kernel and copied the 
resulting /drivers/scsi/aacraid/ directory into the redhat source tree 
for 2.4.20-20.9.  After a dep and clean the build complained of a 
missing compat.h.  I didn't look to see if it was really used or just a 
Makefile dependancy - instead I just copied compat.h from the aacraid 
build 2166 directory.  It built fine.
	I'm now up and running for 35 minutes - longer than any test yet.
	I've looked at the patches vs the mainline kernel, and while the FIBS 
change from 578 to 512 is the only hardware related change, the linked 
list handling has been completely replaced.  The mainline driver and the 
RedHat driver both use the kernel implementation of a double linked 
list.  The new version uses a simple single linked list.  If you're not 
protecting access to this list correctly a double linked list will give 
you at least 2x greater chance (usually more) of a really bad situation. 
I believe it's possible in a single linked list to get away with a 
number of operations without locking access to the list, especially if 
they end up being atomic ops.  I don't think it's possible at all on a 
double linked list.
	So, I'd search in this direction.  I'll send an update this evening if 
things are still running fine, and I'll send one immediately if things 
blow up again.

Andrew

Salyzyn, Mark wrote:

> I have not been able to duplicate this issue, so I am somewhat of a JAFO,
> and am *not* a definitive resource.
> 
> This issue is not just one problem. noapic kernel option and turning off
> HyperThreading have resolved some of the reported issues. Driver changes
> thus far can not eliminate the problem, but can delay the inevitable. Build
> 3157 of the Firmware appears to work fine, Build 3170 fails, but only with
> certain Seagate 15K rpm U320 drives. 
> 
> I may be wrong ... any corrections to my assumptions above would be greatly
> appreciated.
> 
> Sincerely -- Mark Salyzyn
> 
> -----Original Message-----
> From: Thomas Petersen [mailto:tomp at securityminded.net]
> Sent: Tuesday, October 07, 2003 8:52 PM
> To: 'Andrew Mann'
> Cc: linux-poweredge at dell.com; Salyzyn, Mark
> Subject: RE: RedHat 9 aacraid - system fails under extreme disk IO -
> Reproducable test case
> 
> 
> I am pretty disappointed in Dell for failing to follow up on this and
> resolve the issue once and for all.  This is not a new problem but it is
> Dell's responsibility to rectify it as they -certify- Redhat on the 2650 --
> regardless if it's a hardware or software issue Dell is responsible to their
> customers.  
> 
> If this was an issue on the Microsoft platform you can bet Dell would of
> worked with Microsoft and issued a patch/update long before it became a wide
> spread problem.  I have always been a huge fan of Dell equipment but their
> failure in this instance to support what they sell is very troubling. 
> 
> Don't get me wrong I will probably purchase Dell servers again in the future
> (though not the 2650) but can anyone name one problem affecting the
> Microsoft platform, related to Dell hardware and had a problem of this
> magnitude, that went unresolved for as long as this one has?  System lockups
> are -totally- unacceptable.  
> 
> I guess when people start choosing with their checkbooks Dell might wake up.
> 
> Thomas Petersen
> SecurityMinded Technologies 
> 
> 
>>>-----Original Message-----
>>>From: Andrew Mann [mailto:amann at mythicentertainment.com] 
>>>Sent: Tuesday, October 07, 2003 6:20 PM
>>>To: linux-poweredge at dell.com
>>>Cc: mark_salyzyn at adaptec.com
>>>Subject: Re: RedHat 9 aacraid - system fails under extreme 
>>>disk IO - Reproducable test case
>>>
>>>
>>>	Unfortunately we've got a good number of 2550s and 
>>>2650s in use, and 
>>>replacing the RAID cards isn't ideal.  Mostly we don't have 
>>>enough load 
>>>to cause this problem, but every now and then we do get an 
>>>unexplained 
>>>lockup that pulls someone out of bed at 2 AM.
>>>	I searched back through the reports of this and found 
>>>some posts from 
>>>Mark Salyzyn referencing AAC_NUM_FIB and AAC_NUM_IO_FIB 
>>>settings.  The 
>>>last comment I see is on 9/9/2003:
>>>"I am suggesting that this value be (AAC_NUM_IO_FIB+64), and 
>>>limited to 
>>>below 512 (the maximum number of hardware FIBS the Firmware 
>>>can absorb). 
>>>I will begin testing the stability and side effects of this input."
>>>	However, I don't see any followup, nor does the latest 
>>>patchset to the 
>>>2.4 series seem to contain any modifications in this area (or 
>>>2.5 or 2.6 
>>>since June 2003).
>>>	Additionally, I've just rebuilt the aacraid module here 
>>
>>>from the RedHat 
>>
>>>SRPM of 2.4.20-20.9 with AAC_NUM_FIB=512 and 
>>>AAC_NUM_IO_FIB=448, rebuilt 
>>>the rdimage and such and got another crash within 5 minutes 
>>>of starting 
>>>the test.
>>>
>>>	I also see a note from Mark on 8/27/2003:
>>>-----
>>>There is code that does the following in the driver:
>>>
>>>	scsicmd->result = DID_OK << 16 | COMMAND_COMPLETE << 8 
>>>| SAM_STAT_TASK_SET_FULL;
>>>	aac_io_done(scsicmd);
>>>	return -1;
>>>
>>>This is *wrong*, because the none zero return causes the 
>>>system to hold 
>>>the command in the queue due to the use of the new error 
>>>handler, yet we 
>>>have also completed the command as `BUSY' *and* as a result of the 
>>>constraints of the aac_io_done call which relocks (on 
>>>io_request_lock) 
>>>the caller had to unlock leaving a hole that SMP machines fill. By 
>>>dropping the result and done calls in these situations, and 
>>>holding the 
>>>locks in the caller of such routines, I believe we will close 
>>>this hole.
>>>
>>>....
>>>
>>>I will report back on my tests of these changes, but will need a 
>>>volunteer with kernel compile experience to report on the success in 
>>>resolving this issue in the field *please*.
>>>-----
>>>
>>>	I'm not familiar enough with the aacraid driver or scsi 
>>>in general to 
>>>gather the code changes necessary.  There also don't appear to be any 
>>>followups.
>>>
>>>	Mark, do you have any updates on this?  I can make code 
>>>changes, 
>>>recompile, and run a test case that reliably reveals the 
>>>problem here if 
>>>that's helpful.
>>>
>>>
>>>I can't see the full panic message, but the parts I can see are 
>>>basically (copied by hand):
>>>
>>>CPU 1
>>>EFLAGS: 00010086
>>>
>>>EIP is at rmqueue [kernel] 0x127  (2.4.20-20.9smp)
>>>eax: c0343400    ebx: c03445dc    ecx: 00000000
>>>edx: b6d7ca63    esi: 00000000    edi: c03445d0
>>>ebp: 00038000    esp: ee643e80     ds: 0068
>>>es: 0068  ss: 0068
>>>
>>>Process dd (pid: 956, stack page = ee643000)
>>>
>>>Call trace:   wakeup_kswapd   0xfb (0xee643e90)
>>>              __aloc_pages_limit   0x57
>>>              __alloc_pages        0x101
>>>              generic_file_write   0x394
>>>              ext3_file_write      0x39
>>>              sys_write            0x97
>>>              system_call          0x33
>>>
>>>	Although aacraid isn't directly implicated here, I can 
>>>reproduce this 
>>>on the 2550s and 2650s (aacraid) but not 1750s (megaraid).
>>>
>>>Andrew
>>>
>>>Paul Anderson wrote:
>>>
>>>
>>>>We had this same issue with our 2650's running AS 2.1.  Don't know 
>>>>that this is the best answer, but it is the one that worked for 
>>>>us...Replace the on board adapter with a PERC 3/DC (LSI) adapter.  
>>>>Make sure that you put it on its own bus, we used slot 
>>>
>>>three.  In 2 of 
>>>
>>>>our 2650's we are even running this with the HBA's for SAN 
>>>>connectivity.  That said, our solution is about 2 weeks 
>>>
>>>old, though I 
>>>
>>>>did run similar tests on the systems after the new install 
>>>
>>>for 8 days 
>>>
>>>>and was unable to make them crash.
>>>>
>>>>Paul
>>>>
>>>>-----Original Message-----
>>>>From: Andrew Mann [mailto:amann at mythicentertainment.com]
>>>>Sent: Tuesday, October 07, 2003 12:47 PM
>>>>To: linux-poweredge at dell.com
>>>>Cc: Matt Domsch; deanna_bonds at adaptec.com; alan at redhat.com
>>>>Subject: RedHat 9 aacraid - system fails under extreme disk IO - 
>>>>Reproducable test case
>>>>
>>>>
>>>>	This has been brought up on the Dell Linux Poweredge 
>>>
>>>list previously,
>>>
>>>>but it doesn't appear that a definative solution or reproducable 
>>>>situation has been presented.  It also seems like the 
>>>
>>>previous reports 
>>>
>>>>involved both heavy disk IO as well as heavy network 
>>>
>>>traffic, and so the 
>>>
>>>>NIC driver was suspect.
>>>>	Since we have a number of 2550s and 2650s using the 
>>>
>>>onboard PERC3/Di 
>>>
>>>>raid controller (aacraid driver), this issue concerns us.
>>>>
>>>>	The following script was run with 6 instances at once 
>>>
>>>on two 2550s 
>>>
>>>>and
>>>>one 2650.
>>>>
>>>>2550 configuration
>>>>2 x P3 1.2 Ghz  kernel: 2.4.20-20.9smp #1 SMP
>>>>1GB of ram, 2GB of swap, 2 x 18 GB drives in a raid 1 configuration
>>>>
>>>>2650 configuration
>>>>2 x Xeon 2.2 Ghz   kernel: 2.4.20-20.9smp #1 SMP
>>>>2GB of ram, 2GB of swap, 2 x 18 GB drives in a raid 1 configuration 
>>>>Hyperthreading enabled
>>>>
>>>>
>>>>	The 2550s fail within 30 minutes of starting the tests 
>>>
>>>each time 
>>>
>>>>(tests
>>>>were run 6 times in a row).  The 2650 failed prior to 2.5 
>>>
>>>days (only 1 
>>>
>>>>test run due to duration before failure).  In some cases the 2550 
>>>>displayed a null pointer dereference in the kernel.  I'll copy down 
>>>>details next time I can catch it on screen.  It does not 
>>>
>>>get logged to 
>>>
>>>>disk, which doesn't surprise me in this situation.  In most 
>>>
>>>cases the 
>>>
>>>>screen was blank (due to APM I'd guess?).
>>>>	The systems still respond to pings, but do not respond 
>>>
>>>to keyboard 
>>>
>>>>actions and do not complete any tcp connections.  These 
>>>
>>>systems do not 
>>>
>>>>have a graphical desktop installed, and in fact have a 
>>>
>>>fairly minimal 
>>>
>>>>set of packages installed at all.
>>>>	I don't know why the 2550 would consistantly fail in 
>>>
>>>such a brief 
>>>
>>>>period while the 2650 would take a much longer time before failure. 
>>>>I've been running the same tests on a 1750 (PERC4/Di - 
>>>
>>>Megaraid based) 
>>>
>>>>for some days now without a failure.
>>>>	I plan on testing a non-SMP kernel on the 2550 next - 
>>>
>>>not because we 
>>>
>>>>can run things that way, but to maybe give some more clues.
>>>>
>>>>	The following script creates a 300 MB file, then rm's 
>>>
>>>it, then does 
>>>
>>>>it
>>>>all over again.  For my tests I ran 6 of these concurrently.  Don't 
>>>>expect the system to respond to much while these are 
>>>
>>>running, though I 
>>>
>>>>was able to get decent updates from top.
>>>>	Alter the script as you see fit, I'm no guru with bash 
>>>
>>>scripting!
>>>
>>>>cat diskgrind.sh
>>>>#!/bin/sh
>>>>
>>>>
>>>>MEGS=300
>>>>TOTAL=0
>>>>
>>>>while [ "1" != "0" ]; do
>>>>         dd ibs=1048576 count=$MEGS if=/dev/zero 
>>>
>>>of=/test/diskgrind.$$
>>>
>>>>2>&1 | cat >/dev/null
>>>>         rm -f /test/diskgrind.$$
>>>>         TOTAL=`expr $TOTAL + $MEGS`
>>>>         echo "[$$] Completed $TOTAL megs."
>>>>done
>>>>
>>>>
>>>>./diskgrind.sh &
>>>>./diskgrind.sh &
>>>>./diskgrind.sh &
>>>>./diskgrind.sh &
>>>>./diskgrind.sh &
>>>>./diskgrind.sh &
>>>>
>>>>
>>>>
>>>>Andrew
>>>>
>>>
>>>-- 
>>>Andrew Mann
>>>Systems Administrator
>>>Mythic Entertainment
>>>703-934-0446 x 224
>>>
>>>_______________________________________________
>>>Linux-PowerEdge mailing list
>>>Linux-PowerEdge at dell.com 
>>>
>>>>>http://lists.us.dell.com/mailman/listinfo/linux->>poweredge
>>>
>>>
>>>Please read the FAQ at 
>>>http://lists.us.dell.com/faq or search the list archives at 
> 
> http://lists.us.dell.com/htdig/
> 
> 

-- 
Andrew Mann
Systems Administrator
Mythic Entertainment
703-934-0446 x 224




More information about the Linux-PowerEdge mailing list