Am So., 16. Mai 2021 um 21:36 Uhr schrieb Ilya Dryomov <idryomov@gmail.com>:
On Sun, May 16, 2021 at 8:06 PM Markus Kienast <mark@trickkiste.at> wrote:
Am So., 16. Mai 2021 um 19:38 Uhr schrieb Ilya Dryomov <
On Sun, May 16, 2021 at 4:18 PM Markus Kienast <mark@trickkiste.at>
wrote:
Am So., 16. Mai 2021 um 15:36 Uhr schrieb Ilya Dryomov <
idryomov@gmail.com>:
On Sun, May 16, 2021 at 12:54 PM Markus Kienast <mark@trickkiste.at>
wrote:
Hi Ilya,
unfortunately I can not find any "missing primary copy of ..."
error in the logs of my 3 OSDs.
The NVME disks are also brand new and there is not much traffic on
The only error keyword I find are those two messages in osd.0 and
osd.1 logs shown below.
BTW the error posted before actually concerns osd1. The one I
Hi Markus,
Please don't ever paste log messages from other bug reports again. Your email said "I am seeing these messages ..." and I spent a fair amount of time staring at the code trying to understand how an issue that was fixed several releases ago could resurface.
The numbers in the log message mean specific things. For example it is immediately obvious that
get_reply osd1 tid 11 data 4164 > preallocated 4096, skipping
is not related to
get_reply osd2 tid 1459933 data 3248128 > preallocated 131072,
skipping
even though they probably look the same to you.
Sorry, I was not aware of that.
[ 10.331119] libceph: mon1 (1)10.101.0.27:6789 session
established
[ 10.331799] libceph: client175444 fsid b0f4a188-bd81-11ea-8849-97abe2843f29 [ 10.336866] libceph: mon0 (1)10.101.0.25:6789 session established [ 10.337598] libceph: client175444 fsid b0f4a188-bd81-11ea-8849-97abe2843f29 [ 10.349380] libceph: get_reply osd1 tid 11 data 4164 >
4096, skipping
Please paste the entire boot log and "rbd info" output for the affected image.
elias@maas:~$ rbd info squashfs/ltsp-01 rbd image 'ltsp-01': size 3.5 GiB in 896 objects order 22 (4 MiB objects) snapshot_count: 0 id: 23faade1714 block_name_prefix: rbd_data.23faade1714 format: 2 features: layering, exclusive-lock, object-map, fast-diff, deep-flatten op_features: flags: create_timestamp: Mon Jan 11 12:09:22 2021 access_timestamp: Wed Feb 24 10:55:17 2021 modify_timestamp: Mon Jan 11 12:09:22 2021
I don't have the boot log available right now, but you can watch a video of the boot process right here: https://photos.app.goo.gl/S8PssYu2VAr4CSeg7
It seems to be consistently "tid 11" consistently, while in this video it was "data 4288" not "data 4164" as above. But the image has been modified in the meantime, as far as I can recall, so that might be due to
elias@maas:~$ juju ssh ceph-osd/2 sudo zgrep -i error
/var/log/ceph/ceph-osd.0.log
2021-05-16T08:52:56.872+0000 7f0b262c2d80 4 rocksdb: Options.error_if_exists: 0 2021-05-16T08:52:59.872+0000 7f0b262c2d80 4 rocksdb: Options.error_if_exists: 0 2021-05-16T08:53:00.884+0000 7f0b262c2d80 1 osd.0 8599 warning: got an error loading one or more classes: (1) Operation not permitted
elias@maas:~$ juju ssh ceph-osd/0 sudo zgrep -i error /var/log/ceph/ceph-osd.1.log 2021-05-16T08:49:52.971+0000 7fb6aa68ed80 4 rocksdb: Options.error_if_exists: 0 2021-05-16T08:49:55.979+0000 7fb6aa68ed80 4 rocksdb: Options.error_if_exists: 0 2021-05-16T08:49:56.828+0000 7fb6aa68ed80 1 osd.1 8589 warning: got an error loading one or more classes: (1) Operation not permitted
How can I find our more about this bug? It keeps coming back every two weeks and I need to restart all OSDs to make it go away for another two weeks. Can I check "tid 11 data 4164" somehow. I find no documentation, what a tid actually is and how I could perform a read test on it.
So *just* restarting the three OSDs you have makes it go away?
What is meant by restarting? Rebooting the node or simply restarting the OSD process?
I did reboot all OSD nodes and since the MON and FS nodes run as LXD/juju instances on them, they were rebooted as well.
Another interesting detail is, that the problem does only seem to
affect booting up from this RBD but not operation per se. The thin clients already booted from this RBD continue working.
I take it that the affected image is mapped on multiple nodes? If
so,
on how many?
Currently "squashfs/ltsp-01" is mapped on 4 nodes. As the pool name indicates, the FS was converted to squashfs and is
idryomov@gmail.com>: them. posted was copied from somebody elses bug report, which had similar errors. Here are my original error messages on LTSP boot: preallocated that reason. therefore mounted read-only, while the underlying dev might actually not be mounted read-only, as there does not seem to be an option available to mount RO via /sys/bus/rbd/add_single_major or /sys/bus/rbd/add.
As far as I can tell, the only way to force RO is to map a snapshot
instead.
Are you writing to /sys/bus/rbd/add_single_major directly instead of using the rbd tool?
Yes. Line 110 https://github.com/trickkiste/ltsp/blob/feature-boot_method-rbd/debian/ltsp-...
echo "${mons} name=${user},secret=${key} ${pool} ${image} ${snap}" > ${rbd_bus}
All systems run: Ubuntu 20.04.2 LTS Kernel 5.8.0-53-generic ceph version 15.2.8 (bdf3eebcd22d7d0b3dd4d5501bee5bac354d5b55)
octopus (stable)
The cluster has been setup with Ubuntu MAAS/juju, consists of * 1 MAAS server * with 1 virtual LXD juju controller * 3 OSD servers with one 2 TB Nvme SSD each for ceph and a 256
SATA SSD for the operating system.
* each OSD contains a virtualized LXD MON and an LXD FS server (setup through juju, see juju yaml file attached).
Can you describe the client side a bit more? How many clients do you have? How many of them are active at the same time?
Currently, there are only 4 active clients but the system is intended to being able to sustain 100s of clients. We are using an RBD as boot device for PXE booted thin clients, you might have heard of the Linux Terminal Server Project (ltsp.org). We adapted the stack to support booting from RBD.
How many active clients there were at the time when the image couldn't be mapped? I suspect between 60 and 70?
No, just 4. Most of the time 3 still running and working correctly and one stuck at reboot.
Maybe the sum of all LTSP client reboots since I cleared the problem by rebooting the OSDs could amount to 60-70. I do not know, as we are not logging that currently.
The next time it happens, check the output of "rbd status" for that image. If you see around 65 watchers, that is it. With exclusive-lock feature enabled on the image, the current kernel implementation can't handle more than that.
OK, currently I am seeing 5, which is one more than the number of clients we have. So it seems these watchers do not timeout after reboot or hard reset.
Is there any way to make these watchers time out?
They are supposed to time out after 30 seconds. Does the IP address of the rogue watch offer a clue?
Not really, I see 3 watchers for one of my client IPs right now and one for each of the others. Have to investigate further. The clients are assigned the same IP on each bootup, so shouldn't the watchers either time out or be "claimed/taken over" by the newly booted client?
Note that when the mapping gets stuck on that preallocated check, it still maintains the watch so it's not going to time out in that case.
I will investigate, if it is only one client, producing all the watchers until the limit is reached, or if all clients produce this problem together and then report back on the issue.
Watches are established if the image is mapped read-write. For your squashfs + overlayfs use case, it's not only better to map read-only just in case, you actually *need* to do that to avoid watches being established.
If you are writing to /sys/bus/rbd/add_single_major directly, append "ro" somewhere in the options part of the string:
ip:port,... name=myuser,secret=mysecret rbd ltsp-01 - # read-write
ip:port,... name=myuser,secret=mysecret,ro rbd ltsp-01 - # read-only
Thank you, we will add this missing piece to our rbd initrd code.
Are you a ceph dev? Could you make sure to add this to kernel documentation too! https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-bus-rbd
Map options are documented in the rbd man page:
https://docs.ceph.com/en/latest/man/8/rbd/#kernel-rbd-krbd-options
Thanks for the tip! However, I would argue that these krbd options should either be available in the kernel documentation as well, or at least this man page should be mentioned in the kernel documentation.
There is no mention of that option currently. I might even have tried this but it might not have worked. Not sure,
this has been over a year back.
Also missing in the documentation is, how one could mount a CephFS on
boot!!!
Do you mean booting *from* CephFS, i.e. using it as a root filesystem? Because mounting CephFS on boot after root filesystem is mounted is done through /etc/fstab, like you would mount any other filesystem whether local or network.
Yes, CephFS as a roof filesystem. The reason, why this would be convenient is, that we could then have one VM or client mounted this CephFS in RW mode, therefore being able to make updates to the packages installed etc. while the other clients would be restricted to read-only rights via authx. We would still be able to create snapshots of the last properly working OS configuration but without the additional hassle. Also, if we do OS package upgrades, these upgrades would be instantly available on the clients as well without the need to reboot, since due to the network-fs nature of CephFS all clients would be aware of the underlying changes to the FS.
We are thinking about switching to booting a CephFS in the future. But I would not have any idea and did not find any documentation on how we would approach that - which boot kernel option to use, which sysfs interface could be used, or which tools we must include in initrd.
Generally it would be great if you could include the proper initrd code for RBD and CephFS root filesystems to the Ceph project. You can happily use my code as a starting point.
https://github.com/trickkiste/ltsp/blob/feature-boot_method-rbd/debian/ltsp-...
I think booting from CephFS would require kernel patches. It looks like NFS and CIFS are the only network filesystems supported by the init/root infrastructure in the kernel.
As long as we can do it similar to the way we do with RBD now, that would be sufficient. The actual mapping and mounting in both NFS and RBD does take place in the initrd scripts. The RBD initrd script actually is derived from the NFS initrd script, so I would presume that one could make CephFS mount work in a similar fashion. Question being, which tools are needed to make this work? I guess "mount" provides everything needed already, does it not?
Thank you so much for you support, Markus Thanks,
Ilya