Re: 2 errors when scrubbing - but I don't know what they mean

[Sebastian Ochmann <ochmann@informatik.uni-bonn.de>]

Hello,

 > However, if you find such superblocks checksum mismatch very often
 > during scrub, it maybe
 > there are something wrong with disk!

I'm sorry, but I don't think there's a problem with my disks because I 
was able to trigger the errors that increment the "gen" error counter 
during scrub on a completely different machine and drive today. I 
basically performed some I/O operations on a drive and scrubbed at the 
same time over and over again until I actually saw "super" errors during 
scrub. But the error is reeally hard to trigger. It seems to me like a 
race condition somewhere.

So I went a step further and tried to create a repro for this. It seems 
like I can trigger the errors now once every few minutes with the method 
described below, but sometimes it really takes a long time until the 
error pops up, so be patient when trying this...

For the repro:

I'm using a btrfs image in RAM for this for two reasons: I can scrub 
quickly over and over again and I can rule our hard drive errors. My 
machine has 32 GB of RAM, so that comes in handy here - if you try this 
on a physical drive, make sure to adjust some parameters, if necessary.

Create a tmpfs and a testing image, format as btrfs:

$ mkdir btrfstest
$ cd btrfstest/
$ mkdir tmp
$ mount -t tmpfs -o size=20G none tmp
$ dd if=/dev/zero of=tmp/vol bs=1G count=19
$ mkfs.btrfs tmp/vol
$ mkdir mnt
$ mount -o commit=1 tmp/vol mnt

Note the "commit=1" mount option. It's not strictly necessary, but I 
have the feeling it helps with triggering the problem...

So now we have a 19 GB btrfs filesystem in RAM, mounted in "mnt". What I 
did for performing some artificial I/O operations is to rm and cp a 
linux source tree over and over again. Suppose you have an unpacked 
linux source tree available in the "/somewhere/linux" directory (and 
you're using bash). We'll spawn some loops that keep the filesystem busy:

$ while true; do rm -fr mnt/a; sleep 1.0; cp -R /somewhere/linux mnt/a; 
sleep 1.0; done
$ while true; do rm -fr mnt/b; sleep 1.1; cp -R /somewhere/linux mnt/b; 
sleep 1.1; done
$ while true; do rm -fr mnt/c; sleep 1.2; cp -R /somewhere/linux mnt/c; 
sleep 1.2; done

Now that the filesystem is busy, we'll also scrub it repeatedly (without 
backgrounding, -B):

$ while true; do btrfs scrub start -B mnt; sleep 0.5; done

On my machine and in RAM, each scrub takes 0-1 second and the "total 
bytes scrubbed" should fluctuate (seems to be especially true with 
commit=1, but not sure). Get a beverage of your choice and wait.

(about 10 minutes later)

When I was writing this repro it took about 10 minutes until scrub said:

   total bytes scrubbed: 1.20GB with 2 errors
   error details: super=2
   corrected errors: 0, uncorrectable errors: 0, unverified errors: 0

and in dmesg:

   [15282.155170] btrfs: bdev /dev/loop0 errs: wr 0, rd 0, flush 0, 
corrupt 0, gen 1
   [15282.155176] btrfs: bdev /dev/loop0 errs: wr 0, rd 0, flush 0, 
corrupt 0, gen 2

After that, scrub is happy again and will continue normally until the 
same errors happen again after a few hundred scrubs or so.

So all in all, the error can be triggered using normal I/O operations 
and scrubbing at the right moments, it seems. Even with a btrfs image in 
RAM, so no hard drive error is possible.

Hope anyone can reproduce this and maybe debug it.

Best regards
Sebastian