Re: question about raidframe getting stuck
snip Almost every RAID system out there handles the sudden removal of a disk from the system pretty well. Why? Because it's EASY to create that failure mode. Problem is, in 25 years in this business, I don't recall having seen a hard disk fall out of a computer as a mode of actual failure (I did see a SCSI HBA fall out of a machine once, but that's a different story). snip I had seen that disk-suddenly-out-of-computer failure once. Coincidently enough, it was an OpenBSD system configured only for NAT, about 6 years ago. The IDE hard disk failed sometime at night. When we arrived on the next day at office. Everything was working flawlessly until someone ssh'ed to that machine. My guess is something has gone awry when the syslog went to write that new connection and suddenly the OS discovered that was no HD present. Surprisingly enough, the onboard IDE controller survived, but after installing the new disk, we found the parallel IDE cable faulty and it had to be replaced also. It was not a RAID system though... snip
Re: question about raidframe getting stuck
Hi Nick, I highly appreciate your detailed report about your experiences with RAID systems. That was cool. Surely I don't expect any miracles from RAID anymore. The current plan is to move to a ramdisk based system to get rid of disk access afap, and to use carp to setup a fallback host. Logging is done (non-blocking, hopefully) via network. Many thanx Harri
Re: question about raidframe getting stuck
Stuart Henderson wrote: With IDE (Integrated Drive Electronics), the controller is *on the drive*. A failing drive/controller can do all sorts of nasty things to the host system. So you mean I should not use IDE disks (PATA or SATA), because Raidframe cannot support a failsafe operation with these disks? Regards Harri
Re: question about raidframe getting stuck
Harald Dunkel wrote: Stuart Henderson wrote: With IDE (Integrated Drive Electronics), the controller is *on the drive*. A failing drive/controller can do all sorts of nasty things to the host system. So you mean I should not use IDE disks (PATA or SATA), because Raidframe cannot support a failsafe operation with these disks? rant Your basic assumption that RAID=minimal down time is flawed. The most horrible down times and events you ever see will often involve RAID. That goes for hardware RAID, software RAID, whatever. Almost every RAID system out there handles the sudden removal of a disk from the system pretty well. Why? Because it's EASY to create that failure mode. Problem is, in 25 years in this business, I don't recall having seen a hard disk fall out of a computer as a mode of actual failure (I did see a SCSI HBA fall out of a machine once, but that's a different story). My preferred way to test RAID systems involves a powder-actuated nail gun driving a nail through the platter. Not overly realistic either, but arguably more so than having the drive suddenly being pulled out. The disks get expensive, though. Back to your situation... The drive reports a failure, but not one so horrible that the OS doesn't attempt a retry. So, at what point does the OS just shut down the drive and say, not worth the trouble? If you are running a single drive, you generally want to keep trying as long as there is the slightest hope (another digression: back in the MSDOS v2 days, I had a machine blow a disk such that if I kept hitting Retry enough times, each sector would ultimately be read successfully. Wedged a pen in between the 'R' key and the monitor, went to dinner, and when I came back, I had all my data successfully copied to another drive). In your case, however, you have a drive saying, I'm getting better when you are saying, It'll be stone dead in a moment. You want the OS to whack the drive and toss it on the cart..er..remove it from the RAID set at the first sign of trouble, but that's not a universal answer. Curiously, I've had servers that caused problems BOTH Ways. One kept a drive on-line even though it was having serious problems and should have been declared dead. In a several other cases, the drives reported minor errors and were popped off-line and out of the array when there was really nothing significant wrong with the disks, but the local staff didn't recognize that...and if the right two popped off-line, down went the array. oh, btw: those were both HW RAID. You can run into these problems no matter what you are using. The try too long ones were SATA, the give up too early ones were SCSI. We had 20 servers with the SATA HW mirroring, not a single one lost data, though one got Really Slow until we figured out it was a drive problem. We had 15 SCSI systems which cost about four times as much as the SATA systems..three of those lost data. Complexity kills. Proper operation can be neat. Failures rarely rarely are. There are usually more ways a system can fail than there are ways it can work. It is also really hard to have the drives fail in realistic ways when the designers are watching, and it is really hard to fail something the same exact way again to work out every bug. In your case, you have firewalls, which can be made completely redundant, rather than just making the disk system more complex. Why run RAID in the firewalls, when you can just run CARP and have much more redundancy? Of course, you can have similar problems with CARP, too..we managed to install a proxy package in a CARP'd pair of FWs and didn't notice how fast it filled the disk. One box quit passing packets when the proxy couldn't log any longer, but CARP didn't see the box or interfaces or links as actually failing, so it didn't switch over to the standby system. Happened when both our administrators were out that morning (of course), so when they called me at home, I asked a few questions, and had them hit the power switch on the primary firewall, which instantly got data moving again through the secondary. Consider RAID to be a rapid repair tool, don't expect it to never let you go down (and that's assuming you know how to recover when it actually does fail...and most people I've seen just assume magic happens or they hope they got a job elsewhere by that point). And don't expect to get less down time out of a very complex system compared to a simple system. In particular, when an IDE disk fails, it often does seem to the computer than an entire controller fell out of the system, so don't expect an IDE system to stay up after a drive failure. On the other hand, if you haven't seen a SCSI disk take out an entire SCSI bus, just wait, do enough, you will. Don't expect them to stay up, either. SATA? Ask me again in about ten ten years, but so far, I've seen a SATA drive toss a dead short across the power supply, killing the RAID box, the PS in the computer and the PS on a second
Re: question about raidframe getting stuck
On Thu, Aug 07, 2008 at 12:33:40PM +0200, [EMAIL PROTECTED] wrote: Seems we have some misunderstanding here. I am talking about future events. Of course I don't know in advance which disk fails when. If a disk dies, then its the job of raidframe to detect this event, to mark the disk as bad, and to provide the basic service with the remaining disks, as far as possible. And yet the machine became unresponsive for 30 minutes. This took much too long. Couldn't it be related to the IDE bus? What for noise can a deffective disk on an IDE controller generate when it is failling. I would suggest to you to use SCSI controllers and disks with hot-swappable functionnalities. -- Olivier Cherrier - Symacx.com mailto:[EMAIL PROTECTED]
Re: question about raidframe getting stuck
On 2008-08-08, Olivier Cherrier [EMAIL PROTECTED] wrote: Couldn't it be related to the IDE bus? What for noise can a deffective disk on an IDE controller generate when it is failling. With IDE (Integrated Drive Electronics), the controller is *on the drive*. A failing drive/controller can do all sorts of nasty things to the host system.
Re: question about raidframe getting stuck
On Thu, Aug 07, 2008 at 09:27:24AM +0200, Harald Dunkel wrote: I've got a configuration issue with Raidframe: Our gateway/firewall runs a raid1 for the system disk. No swap partition. Recently one of the raid disks (wd0) showed some problem: Aug 2 17:22:35 fw01 /bsd: wd0(pciide0:0:0): timeout Aug 2 17:53:52 fw01 /bsd: type: ata Aug 2 17:53:52 fw01 /bsd: c_bcount: 16384 Aug 2 17:53:52 fw01 /bsd: c_skip: 0 Aug 2 17:53:52 fw01 /bsd: pciide0:0:0: bus-master DMA error: missing interrupt, status=0x21 Aug 2 17:53:52 fw01 /bsd: pciide0 channel 0: reset failed for drive 0 Aug 2 17:53:52 fw01 /bsd: wd0d: device timeout writing fsbn 46172704 of 46172704-46172735 (wd0 bn 50368000; cn 49968 tn 4 sn 4), retrying : : Aug 2 17:53:52 fw01 /bsd: wd0d: device timeout writing fsbn 46172704 of 46172704-46172735 (wd0 bn 50368000; cn 49968 tn 4 sn 4) Aug 2 17:53:52 fw01 /bsd: raid0: IO Error. Marking /dev/wd0d as failed. Aug 2 17:53:52 fw01 /bsd: raid0: node (Wpd) returned fail, rolling forward Aug 2 17:53:52 fw01 /bsd: pciide0:0:0: not ready, st=0xd0BSY,DRDY,DSC, err=0x00 Aug 2 17:53:52 fw01 /bsd: pciide0 channel 0: reset failed for drive 0 Aug 2 17:53:52 fw01 /bsd: wd0d: device timeout writing fsbn 46137472 of 46137472-46137503 (wd0 bn 50332768; cn 49933 tn 4 sn 52), retrying : : Aug 2 17:53:53 fw01 /bsd: pciide0:0:0: not ready, st=0xd0BSY,DRDY,DSC, err=0x00 Aug 2 17:53:53 fw01 /bsd: pciide0 channel 0: reset failed for drive 0 Aug 2 17:53:53 fw01 /bsd: wd0d: device timeout writing fsbn 46152320 of 46152320-46152343 (wd0 bn 50347616; cn 49948 tn 0 sn 32) Aug 2 17:53:53 fw01 /bsd: raid0: node (Wpd) returned fail, rolling forward Surely wd0 is defect. Can happen. But my problem is that the machine became unresponsive for 30 minutes. Even a ping did not work. This is not what I would expect from a raid system. What would you suggest to reduce the waiting time? 2 minutes would be OK, but 30 minutes downtime are a _huge_ problem. Do I have to expect the same for a raid5 built from 9 disks, but with a higher probability, because there are more disks in the loop? Your best bet is to replace the disk. 30 minutes wait time seems a bit odd though. I have a similar situation where one disk is having problems, requiring the disk to restart, but that only takes approx. a minute. You can mark the disk as bad and replace it before the other disk fails I guess (after all, there's not much point in relying on a faulty disk). Ciao, Ariane
Re: question about raidframe getting stuck
Ariane van der Steldt wrote: Your best bet is to replace the disk. 30 minutes wait time seems a bit odd though. I have a similar situation where one disk is having problems, requiring the disk to restart, but that only takes approx. a minute. You can mark the disk as bad and replace it before the other disk fails I guess (after all, there's not much point in relying on a faulty disk). The problem is not replacing the disk, but how to avoid 30 minutes downtime due to some low level kernel routine getting stuck. Regards Harri
Re: question about raidframe getting stuck
On Thu, Aug 07, 2008 at 11:41:59AM +0200, Harald Dunkel wrote: Ariane van der Steldt wrote: Your best bet is to replace the disk. 30 minutes wait time seems a bit odd though. I have a similar situation where one disk is having problems, requiring the disk to restart, but that only takes approx. a minute. You can mark the disk as bad and replace it before the other disk fails I guess (after all, there's not much point in relying on a faulty disk). The problem is not replacing the disk, but how to avoid 30 minutes downtime due to some low level kernel routine getting stuck. Mark it as a bad disk? If you do that, the raid code should do no more requests to the disk. Ariane
Re: question about raidframe getting stuck
Harald Dunkel wrote: Ariane van der Steldt wrote: Your best bet is to replace the disk. 30 minutes wait time seems a bit odd though. I have a similar situation where one disk is having problems, requiring the disk to restart, but that only takes approx. a minute. You can mark the disk as bad and replace it before the other disk fails I guess (after all, there's not much point in relying on a faulty disk). The problem is not replacing the disk, but how to avoid 30 minutes downtime due to some low level kernel routine getting stuck. Regards Harri Presumably this was after a reboot? If so, the trick is to move the 'raidctl -P all' line from /etc/rc to /etc/rc.local and add a '' so it runs as a background process. Regards Noth
Re: question about raidframe getting stuck
nothingness wrote: Presumably this was after a reboot? If so, the trick is to move the 'raidctl -P all' line from /etc/rc to /etc/rc.local and add a '' so it runs as a background process. There was no reboot involved. Before this event the machine was running for weeks, and it is still running. Regards Harri
Re: question about raidframe getting stuck
Ariane van der Steldt wrote: On Thu, Aug 07, 2008 at 11:41:59AM +0200, Harald Dunkel wrote: Ariane van der Steldt wrote: Your best bet is to replace the disk. 30 minutes wait time seems a bit odd though. I have a similar situation where one disk is having problems, requiring the disk to restart, but that only takes approx. a minute. You can mark the disk as bad and replace it before the other disk fails I guess (after all, there's not much point in relying on a faulty disk). The problem is not replacing the disk, but how to avoid 30 minutes downtime due to some low level kernel routine getting stuck. Mark it as a bad disk? If you do that, the raid code should do no more requests to the disk. Seems we have some misunderstanding here. I am talking about future events. Of course I don't know in advance which disk fails when. If a disk dies, then its the job of raidframe to detect this event, to mark the disk as bad, and to provide the basic service with the remaining disks, as far as possible. Looking at the log file it seems that raidframe _did_ mark the disk as bad: : Aug 2 17:53:52 fw01 /bsd: raid0: IO Error. Marking /dev/wd0d as failed. Aug 2 17:53:52 fw01 /bsd: raid0: node (Wpd) returned fail, rolling forward : And yet the machine became unresponsive for 30 minutes. This took much too long. Regards Harri