Date: Fri, 25 Jan 2013 13:32:10 -0800 From: Jeremy Chadwick <jdc@koitsu.org> To: Alexander Motin <mav@FreeBSD.org> Cc: freebsd-fs@freebsd.org, avg@freebsd.org Subject: Re: disk "flipped" - a known problem? Message-ID: <20130125213209.GA1858@icarus.home.lan> In-Reply-To: <20130125212559.GA1772@icarus.home.lan> References: <20130121221617.GA23909@icarus.home.lan> <50FED818.7070704@FreeBSD.org> <20130125083619.GA51096@icarus.home.lan> <20130125211232.GA3037@icarus.home.lan> <20130125212559.GA1772@icarus.home.lan>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, Jan 25, 2013 at 01:25:59PM -0800, Jeremy Chadwick wrote: > On Fri, Jan 25, 2013 at 01:12:32PM -0800, Jeremy Chadwick wrote: > > On Fri, Jan 25, 2013 at 12:36:19AM -0800, Jeremy Chadwick wrote: > > > On Tue, Jan 22, 2013 at 08:19:04PM +0200, Alexander Motin wrote: > > > > On 22.01.2013 00:16, Jeremy Chadwick wrote: > > > > > (Please keep me CC'd as I am not subscribed) > > > > > > > > > > WRT this: > > > > > > > > > > http://lists.freebsd.org/pipermail/freebsd-fs/2013-January/016197.html > > > > > > > > > > I can reproduce the first problem 100% of the time on my home system > > > > > here. I can provide hardware specs if needed, but the important part is > > > > > that I'm using RELENG_9 / r245697, the controller is an ICH9R in AHCI > > > > > mode (and does not share an IRQ), hot-swap bays are in use, and I'm > > > > > using ahci.ko. > > > > > > > > > > I also want to make this clear to Andriy: I'm not saying "there's a > > > > > problem with your disk". In my case, I KNOW there's a problem with the > > > > > disk (that's the entire point to my tests! :-) ). > > > > > > > > > > In my case the disk is a WD Raptor (150GB, circa 2006) that has a very > > > > > badly-designed firmware that goes completely catatonic when encountering > > > > > certain sector-level conditions. That's not the problem though -- the > > > > > problem is with FreeBSD apparently getting confused as to the internal > > > > > state of its devices after a device falls off the bus and comes back. > > > > > Explanation: > > > > > > > > > > 1. System powered off; disk is attached; system powered on, shows up as > > > > > ada5. Can communicate with device in every way (the way I tend to test > > > > > simple I/O is to use "smartctl -a /dev/ada5"). This disk has no > > > > > filesystems or other "stuff" on it -- it's just a raw disk, so I believe > > > > > the g_wither_washer oddity does not apply in this situation. > > > > > > > > > > 2. "dd if=/dev/zero of=/dev/ada5 bs=64k" > > > > > > > > > > 3. Drive hits a bad sector which it cannot remap/deal with. Drive > > > > > firmware design flaw results in drive becoming 100% stuck trying to > > > > > re-read the sector and work out internal decisions to do remapping or > > > > > not. Drive audibly clicking during this time (not actuator arm being > > > > > reset to track 0 noise; some other mechanical issue). Due to firmware > > > > > issue, drive remains in this state indefinitely. > > > > > > > > > > 4. FreeBSD CAM reports repeated WRITE_FPDMA_QUEUED (i.e. writes using NCQ) > > > > > errors every 30 seconds (kern.cam.ada.default_timeout), for a total of 5 > > > > > times (kern.cam.da.retry_count+1). > > > > > > > > > > 5. FreeBSD spits out similar messages you see; retries exhausted, > > > > > cam_periph_alloc error, and devfs claims device removal. > > > > > > > > > > 6. Drive is still catatonic of course. Only way to reset the drive is > > > > > to power-cycle it. Drive removed from hot-swap bay, let sit for 20 > > > > > seconds, then is reinserted. > > > > > > > > > > 7. FreeBSD sees the disk reappear, shows up much like it did during #1, > > > > > except... > > > > > > > > > > 8. "smartctl -a /dev/ada5" claims no such device or unknown device type > > > > > (I forget which). "ls -l /dev/ada5" shows an entry. "camcontrol > > > > > devlist" shows the disk on the bus, yet I/O does not work. If I > > > > > remember right, re-attempting the dd command returns some error (I > > > > > forget which). > > > > > > > > > > 9. "camcontrol rescan all" stalls for quite some time when trying to > > > > > communicate with entry 5, but eventually does return (I think with some > > > > > error). camcontrol reset all" works without a hitch. "camcontrol > > > > > devlist" during this time shows the same disk on ada5 (which to me means > > > > > ATA IDENTIFY, i.e. vendor strings, etc. are reobtained somehow, meaning > > > > > I/O works at some level). > > > > > > > > > > 10. System otherwise works fine, but the only way to bring back > > > > > usability of ada5 is to reboot ("shutdown -r now"). > > > > > > > > > > To me, this looks like FreeBSD at some layer within the kernel (or some > > > > > driver (I don't know which)) is internally confused about the true state > > > > > of things. > > > > > > > > > > Alexander, do you have any ideas? > > > > > > > > > > I can enable CAM debugging (I do use options CAMDEBUG so I can toggle > > > > > this with camcontrol) as well as take notes and do a full step-by-step > > > > > diagnosis (along with relevant kernel output seen during each phase) if > > > > > that would help you. And I can test patches but not against -CURRENT > > > > > (will be a cold day in hell before I run that, sorry). > > > > > > > > Command timeout itself is not a reason for AHCI driver to drop the disk, > > > > neither it is for CAM in case of payload requests. Disk can be dropped > > > > if controller report device absence detected by SATA PHY, or by errors > > > > during device reinitialization after reset by CAM SATA XPT. > > > > > > I have some theories as to why this is happening and it relates to the > > > underlying design of the drive firmware and the drive controller used. > > > I could write some pseudo-code showing how I believe the drive behaves, > > > but it's really besides the point, as you point out below. > > > > > > > What is interesting, is what exactly goes on after disk got stuck and > > > > you have removed it. In normal case controller should immediately report > > > > PHY status change, driver should run PHY reset and see that link is > > > > lost. It should trigger bus rescan for CAM, that should invalidate > > > > device. That should make dd abort with error. After dd gone, device > > > > should be destroyed and ready for reattachment. > > > > > > Yup, that sounds exactly like what should happen. I know that in > > > userland (dd) the command eventually does abort/fail with an error (I > > > believe I/O error or some other message), and that's good. The device > > > disappearing can also be confirmed. It's after the drive is > > > power-cycled (to bring it back online) where its re-tasted and I/O (at > > > the kernel level) works, but now userland utilities interfacing with > > > /dev/ada5 insist "unknown device" or "no such device". It's easier to > > > show than it is to explain. My theory is that there is some kind of > > > internal (kernel-level) "state" that is not being reset correctly when a > > > device is lost and then brought back. > > > > > > > So it should be great if you start with the full verbose dmesg from the > > > > boot up to the moment when system becomes stable after disk removal. If > > > > it won't be enough, we can enable some more debugging with `camcontrol > > > > debug -IPXp BUS`, where BUS is the bus number from `camcontrol devlist`. > > > > > > This is exactly what I needed; thank you! > > > > > > I'll spend some time tomorrow collecting the data + documenting and will > > > provide the results once I've compiled them. This will be more useful > > > than speculation on my part. > > > > Finished. > > > > http://jdc.koitsu.org/freebsd/ahci_cam_testing/ > > > > I was not able to get the cam_periph_alloc error message; I'll talk > > about that at the end of my mail in attempt to stay focused on what I > > did find / what I was able to reproduce. > > > > Things start to get interesting around phase 23. > > > > Phase 31 is where things are confirmed broken in some way ("no such file > > or directory" even though /dev/ada5 is there). > > > > Direct I/O to /dev/ada5 still works (shown in phase 33), but smartctl > > ceases to work ***to that device only*** from then onward (e.g. smartctl > > ada0 works fine). A reboot is needed to recover from this. > > > > I'm aware that smartmontools uses xpt(4), and I think therein is where > > the issue is. The only difference between the tests/phases is that I > > issued "camcontrol reset" and "camcontrol rescan" prior to the breakage. > > > > Based on CAMDEBUG output in phase 36 it looks like xpt(4) is spinning on > > something internally and causing what I'm reporting. I can reproduce it > > reliably at least. > > > > Let me know what else I can do to help. If I need to turn on CAMDEBUG > > and re-do some of the tests + provide full kernel/CAMDEBUG output during > > each phase, no problem, just let me know what you need. I just hate > > risking interspersed kernel printf output... > > I just realised that the numbering scheme I was using for the phases is > *completely* buggered. I obviously made a manual typo at some point and > it just proliferated from there. > > I'll see if I can figure out where my mistake was and clean it up. Yeah, there's no way for me to easily work out what happened here. Sorry, I'm going to have to re-do these tests. I'm still fairly certain the issue has to do with camcontrol reset or camcontrol rescan though, but I can't work out exactly where my typo was in the files. -- | Jeremy Chadwick jdc@koitsu.org | | UNIX Systems Administrator http://jdc.koitsu.org/ | | Mountain View, CA, US | | Making life hard for others since 1977. PGP 4BD6C0CB |
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20130125213209.GA1858>