Re: race condition mounting multi-device iscsi volume, not resolved in newer kernels

[Andrei Borzenkov <arvidjaar@gmail.com>]

On 22.07.2023 09:04, Eric Levy wrote:
> 
> 
> On Sat, Jul 22 2023 at 07:34:26 AM +0300, Andrei Borzenkov
> <arvidjaar@gmail.com> wrote:
> 
>> Log with timestamps could be more useful.
> 
> I would need to generate a report showing millisecond granularity in
> order to reveal intervals not appearing as zero duration.
> 
>> Something tries to mount this filesystem before all its devices are
>> present. You need to find out what does it. btrfs is just a messenger
>> here.
> 
> The systemd mount unit is defined to start after iscsi finishes
> initializing.

"iSCSI finishes initializing" just means iscsid was started. The actual 
device discovery can take arbitrary amount of time due to network 
condition, target behavior and overall system load. So condition "iscsi 
finishes initializing" is absolutely wrong for "we can now start 
mounting remote devices".

With default btrfs udev rules systemd should wait for device required by 
mount unit to become ready; in case of btrfs all devices except the last 
discovered one will be "missing" from systemd point of view and only the 
last device should allow mount unit to start.

Provide full output of

journalctl -b --no-pager --full

immediately after boot (before doing anything else), it may give some hints.

> The idea that the kernel or filesystem layers wer causing
> problems follows from a remark given to me through a different thread
> that I opened just days ago. I have tried many different variations for
> the systemd mount unit without reaching any success for the issue.
> 

Systemd was never designed for dynamic device management when devices 
can come and go at arbitrary time and even less for multi-device 
filesystems.