On 2019-02-08 13:10, waxhead wrote:
Austin S. Hemmelgarn wrote:
On 2019-02-07 13:53, waxhead wrote:
Austin S. Hemmelgarn wrote:
On 2019-02-07 06:04, Stefan K wrote:
Thanks, with degraded as kernel parameter and also ind the fstab
it works like expected
That should be the normal behaviour, cause a server must be up and
running, and I don't care about a device loss, thats why I use a
RAID1. The device-loss problem can I fix later, but its important
that a server is up and running, i got informed at boot time and
also in the logs files that a device is missing, also I see that if
you use a monitoring program.
No, it shouldn't be the default, because:
* Normal desktop users _never_ look at the log files or boot info,
and rarely run monitoring programs, so they as a general rule won't
notice until it's already too late. BTRFS isn't just a server
filesystem, so it needs to be safe for regular users too.
I am willing to argue that whatever you refer to as normal users
don't have a clue how to make a raid1 filesystem, nor do they care
about what underlying filesystem their computer runs. I can't quite
see how a limping system would be worse than a failing system in this
case. Besides "normal" desktop users use Windows anyway, people that
run on penguin powered stuff generally have at least some technical
knowledge.
Once you get into stuff like Arch or Gentoo, yeah, people tend to have
enough technical knowledge to handle this type of thing, but if you're
talking about the big distros like Ubuntu or Fedora, not so much.
Yes, I might be a bit pessimistic here, but that pessimism is based on
personal experience over many years of providing technical support for
people.
Put differently, human nature is to ignore things that aren't
immediately relevant. Kernel logs don't matter until you see
something wrong. Boot messages don't matter unless you happen to see
them while the system is booting (and most people don't). Monitoring
is the only way here, but most people won't invest the time in proper
monitoring until they have problems. Even as a seasoned sysadmin, I
never look at kernel logs until I see any problem, I rarely see boot
messages on most of the systems I manage (because I'm rarely sitting
at the console when they boot up, and when I am I'm usually handling
startup of a dozen or so systems simultaneously after a network-wide
outage), and I only monitor things that I know for certain need to be
monitored.
So what you are saying here is that distro's that use btrfs by default
should be responsible enough to make some monitoring solution if they
allow non-technical users to create a "raid"1 like btrfs filesystem in
the first place. I don't think that many distros install some S.M.A.R.T.
monitoring solution either... in which case you are worse off with a
non-checksumming filesystem.
Actually, more than you probably realize do (Windows does by default
these days, so the big distros that want to compete for desktop users
need to as well), and many have trivial to set up monitoring for MD and
LVM arrays as well.
Since the users you refer to basically ignores the filesystem anyway I
can't see why this would be an argument at all...
My argument here is that we shouldn't assume users will know what
they're doing. It's the same logic behind the saner distros not
defaulting to using BTRFS for installation, if they do and BTRFS causes
the user to lose data, the distro will usually get blamed, even if it
was not at all their fault. Similarly, if a user chooses to use BTRFS
without doing their research, it's very likely that any data loss, even
if it's caused by the user themself not doing things sensibly, will be
blamed on BTRFS.
* It's easily possible to end up mounting degraded by accident if
one of the constituent devices is slow to enumerate, and this can
easily result in a split-brain scenario where all devices have
diverged and the volume can only be repaired by recreating it from
scratch.
Am I wrong or would not the remaining disk have the generation number
bumped on every commit? would it not make sense to ignore
(previously) stale disks and require a manual "re-add" of the failed
disks. From a users perspective with some C coding knowledge this
sounds to me (in principle) like something as quite simple.
E.g. if the superblock UUID match for all devices and one (or more)
devices has a lower generation number than the other(s) then the
disk(s) with the newest generation number should be considered good
and the other disks with a lower generation number should be marked
as failed.
The problem is that if you're defaulting to this behavior, you can
have multiple disks diverge from the base. Imagine, for example, a
system with two devices in a raid1 setup with degraded mounts enabled
by default, and either device randomly taking longer than normal to
enumerate. It's very possible for one boot to have one device delay
during enumeration on one boot, then the other on the next boot, and
if not handled _exactly_ right by the user, this will result in both
devices having a higher generation number than they started with, but
neither one being 'wrong'. It's like trying to merge branches in git
that both have different changes to a binary file, there's no sane way
to handle it without user input.
So why do BTRFS hurry to mount itself even if devices are missing? and
if BTRFS still can mount , why whould it blindly accept a non-existing
disk to take part of the pool?!
It doesn't unless you tell it to., and that behavior is exactly what I'm
arguing against making the default here.
Realistically, we can only safely recover from divergence correctly if
we can prove that all devices are true prior states of the current
highest generation, which is not currently possible to do reliably
because of how BTRFS operates.
So what you are saying is that the generation number does not represent
a true frozen state of the filesystem at that point?
It does _only_ for those devices which were present at the time of the
commit that incremented it.
As an example (don't do this with any BTRFS volume you care about, it
will break it), take a BTRFS volume with two devices configured for
raid1. Mount the volume with only one of the devices present, issue a
single write to it, then unmounted it. Now do the same with only the
other device. Both devices should show the same generation number right
now (but it should be one higher than when you started), but the
generation number on each device refers to a different volume state.
Also, LVM and MD have the exact same issue, it's just not as
significant because they re-add and re-sync missing devices
automatically when they reappear, which makes such split-brain
scenarios much less likely.
Which means marking the entire device as invalid, then re-adding it from
scratch more or less...
Actually, it doesn't.
For LVM and MD, they track what regions of the remaining device have
changed, and sync only those regions when the missing device comes back.
For BTRFS, the same thing happens implicitly because of the COW
structure, and you can manually reproduce similar behavior to LVM or MD
by scrubbing the volume and then using balance with the 'soft' filter to
ensure all the chunks are the correct type.
In both cases though, you still get into trouble if each of the devices
gets used separately from each other before being re-synced (though
BTRFS at least has the decency in that situation to not lose any data,
LVM or MD will just blindly sync whichever mirror they happen to pick
over the others).
* We have _ZERO_ automatic recovery from this situation. This makes
both of the above mentioned issues far more dangerous.
See above, would this not be as simple as auto-deleting disks from
the pool that has a matching UUID and a mismatch for the superblock
generation number? Not exactly a recovery, but the system should be
able to limp along.
* It just plain does not work with most systemd setups, because
systemd will hang waiting on all the devices to appear due to the
fact that they refuse to acknowledge that the only way to correctly
know if a BTRFS volume will mount is to just try and mount it.
As far as I have understood this BTRFS refuses to mount even in
redundant setups without the degraded flag. Why?! This is just plain
useless. If anything the degraded mount option should be replaced
with something like failif=X where X would be anything from 'never'
which should get a 2 disk system up with exclusively raid1 profiles
even if only one device is working. 'always' in case any device is
failed or even 'atrisk' when loss of one more device would keep any
raid chunk profile guarantee. (this get admittedly complex in a multi
disk raid1 setup or when subvolumes perhaps can be mounted with
different "raid" profiles....)
The issue with systemd is that if you pass 'degraded' on most systemd
systems, and devices are missing when the system tries to mount the
volume, systemd won't mount it because it doesn't see all the devices.
It doesn't even _try_ to mount it because it doesn't see all the
devices. Changing to degraded by default won't fix this, because it's
a systemd problem.
The same issue also makes it a serious pain in the arse to recover
degraded BTRFS volumes on systemd systems, because if the volume is
supposed to mount normally on that system, systemd will unmount it if
it doesn't see all the devices, regardless of how it got mounted in
the first place.
Why does systemd concern itself about what devices btrfs consist of.
Please educate me, I am curious.
For the same reason that it concerns itself with what devices make up a
LVM volume or an MD array. In essence, it comes down to a couple of
specific things:
* It is almost always preferable to delay boot-up while waiting for a
missing device to reappear than it is to start using a volume that
depends on it while it's missing. The overall impact on the system from
taking a few seconds longer to boot is generally less than the impact of
having to resync the device when it reappears while the system is still
booting up.
* Systemd allows mounts to not block the system booting while still
allowing certain services to depend on those mounts being active. This
is extremely useful for remote management reasons, and is actually
supported by most service managers these days. Systemd extends this all
the way down the storage stack though, which is even more useful,
because it lets disk failures properly cascade up the storage stack and
translate into the volumes they were part of showing up as degraded (or
getting unmounted if you choose to configure it that way).
IOW, there's a special case with systemd that makes even mounting
BTRFS volumes that have missing devices degraded not work.
Well I use systemd on Debian and have not had that issue. In what
situation does this fail?
At one point, if you tried to manually mount a volume that systemd did
not see all the constituent devices present for, it would get unmounted
almost instantly by systemd itself. This may not be the case anymore,
or it may have been how the distros I've used with systemd on them
happened to behave, but either way it's a pain in the arse when you want
to fix a BTRFS volume.
* Given that new kernels still don't properly generate half-raid1
chunks when a device is missing in a two-device raid1 setup, there's
a very real possibility that users will have trouble recovering
filesystems with old recovery media (IOW, any recovery environment
running a kernel before 4.14 will not mount the volume correctly).
Sometimes you have to break a few eggs to make an omelette right? If
people want to recover their data they should have backups, and if
they are really interested in recovering their data (and don't have
backups) then they will probably find this on the web by searching
anyway...
Backups aren't the type of recovery I'm talking about. I'm talking
about people booting to things like SystemRescueCD to fix system
configuration or do offline maintenance without having to nuke the
system and restore from backups. Such recovery environments often
don't get updated for a _long_ time, and such usage is not atypical as
a first step in trying to fix a broken system in situations where
downtime really is a serious issue.
I would say that if downtime is such a serious issue you have a failover
and a working tested backup.
Generally yes, but restoring a volume completely from scratch is almost
always going to take longer than just fixing what's broken unless it's
_really_ broken. Would you really want to nuke a system and rebuild it
from scratch just because you accidentally pulled out the wrong disk
when hot-swapping drives to rebuild an array?
* You shouldn't be mounting writable and degraded for any reason
other than fixing the volume (or converting it to a single profile
until you can fix it), even aside from the other issues.
Well in my opinion the degraded mount option is counter intuitive.
Unless otherwise asked for the system should mount and work as long
as it can guarantee the data can be read and written somehow
(regardless if any redundancy guarantee is not met). If the user is
willing to accept more or less risk they should configure it!
Again, BTRFS mounting degraded is significantly riskier than LVM or MD
doing the same thing. Most users don't properly research things
(When's the last time you did a complete cost/benefit analysis before
deciding to use a particular piece of software on a system?), and
would not know they were taking on significantly higher risk by using
BTRFS without configuring it to behave safely until it actually caused
them problems, at which point most people would then complain about
the resulting data loss instead of trying to figure out why it
happened and prevent it in the first place. I don't know about you,
but I for one would rather BTRFS have a reputation for being
over-aggressively safe by default than risking users data by default.
Well I don't do cost/benefit analysis since I run free software. I do
however try my best to ensure that whatever software I install don't
cause more drawbacks than benefits.
Which is essentially a CBA. The cost doesn't have to equate to money,
it could be time, or even limitations in what you can do with the system.
I would also like for BTRFS to be over-aggressively safe, but I also
want it to be over-aggressively always running or even limping if that
is what it needs to do.
And you can have it do that, we just prefer not to by default.
