Re: [systemd-devel] [survey] BTRFS_IOC_DEVICES_READY return status

[Goffredo Baroncelli <kreijack@libero.it>]

On 2015-06-15 19:38, Lennart Poettering wrote:
> On Mon, 15.06.15 19:23, Goffredo Baroncelli (kreijack@inwind.it) wrote:
> 
>> On 2015-06-15 12:46, Lennart Poettering wrote:
>>> On Sat, 13.06.15 17:09, Goffredo Baroncelli (kreijack@libero.it) wrote:
>>>
>>>>> Further, the problem will be more intense in this eg. if you use dd
>>>>> and copy device A to device B. After you mount device A, by just
>>>>> providing device B in the above two commands you could let kernel
>>>>> update the device path, again all the IO (since device is mounted)
>>>>> are still going to the device A (not B), but /proc/self/mounts and
>>>>> 'btrfs fi show' shows it as device B (not A).
>>>>>
>>>>> Its a bug. very tricky to fix.
>>>>
>>>> In the past [*] I proposed a mount.btrfs helper . I tried to move the logic outside the kernel.
>>>> I think that the problem is that we try to manage all these cases
>>>> from a device point of view: when a device appears, we register the
>>>> device and we try to mount the filesystem... This works very well
>>>> when there is 1-volume filesystem. For the other cases there is a
>>>> mess between the different layers:
>>>
>>>> - kernel
>>>> - udev/systemd
>>>> - initrd logic
>>>>
>>>> My attempt followed a different idea: the mount helper waits the
>>>> devices if needed, or if it is the case it mounts the filesystem in
>>>> degraded mode. All devices are passed as mount arguments
>>>> (--device=/dev/sdX), there is no a device registration: this avoids
>>>> all these problems.
>>>
>>> Hmm, no. /bin/mount should not block for devices. That's generally
>>> incompatible with how the tool is used, and in particular from
>>> systemd. We would not make use for such a scheme in
>>> systemd. /bin/mount should always be short-running.
>>
>> Apart systemd, which are these incompatibilities ? 
> 
> Well, /bin/mount is not a daemon, and it should not be one.

My helper is not a deamon; you was correct the first time: it blocks until all needed/enough devices are appeared.
Anyway this should not be different from mounting a nfs filesystem. Even in this case the mount helper blocks until the connection happened. The block time is not negligible, even tough not long as a device timeout ... 

> 
>>> I am pretty sure that if such automatic degraded mounting should be
>>> supported, then this should be done with some background storage
>>> daemon that alters the effect of the READY ioctl somehow after the
>>> timeout, and then retriggers the devcies so that systemd takes
>>> note. (or, alternatively: such a scheme could even be implemented all
>>> in kernel, based on some configurable kernel setting...)
>>
>> I recognize that this solution provides the maximum compatibility
>> with the current implementation. However it seems too complex to
>> me. Re-trigging a devices seems to me more a workaround than a
>> solution.
> 
> Well, it's not really ugly. I mean, if the state or properties of a
> device change, then udev should update its information about it, and
> that's done via a retrigger. We do that all the time already, for
> example when an existing loopback device gets a backing file assigned
> or removed. I am pretty sure that loopback case is very close to what
> you want to do here, hence retriggering (either from the kernel side,
> or from userspace), appears like an OK thing to do.

What seems strange to me is that in this case the devices don't have changed their status.
How this problem is managed in the md/dm raid cases ?

> 
>> Could a generator do this job ? I.e. this generator (or storage
>> daemon) waits that all (or enough) devices are appeared, then it
>> creates a .mount unit: do you think that it is doable ?
> 
> systemd generators are a way to extend the systemd unit dep tree with
> units. They are very short running, and are executed only very very
> early at boot. They cannot wait for anything, they don#t have access
> to devices and are not run when they are appear.
> 
> Lennart
> 


-- 
gpg @keyserver.linux.it: Goffredo Baroncelli <kreijackATinwind.it>
Key fingerprint BBF5 1610 0B64 DAC6 5F7D  17B2 0EDA 9B37 8B82 E0B5