[Ocfs2-devel] [PATCH] Revert "ocfs2: mount shared volume without ha stack"

[Heming Zhao]

On Mon, Jun 06, 2022 at 10:08:53AM +0800, heming.zhao--- via Ocfs2-devel wrote:
> Hello Junxiao,
> 
> First of all, let's turn to the same channel to discuss your patch.
> There are two features: 'local mount' & 'nocluster mount'.
> I mistakenly wrote local-mount on some place in previous mails.
> This patch revert commit 912f655d78c5d4, which is related with 'nocluster mount'.
> 
> 
> On 6/5/22 00:19, Junxiao Bi wrote:
> > 
> > 
> >> 在 2022年6月4日，上午1:45，heming.zhao at suse.com 写道：
> >>
> >> Hello Junxiao,
> >>
> >>> On 6/4/22 06:28, Junxiao Bi via Ocfs2-devel wrote:
> >>> This reverts commit 912f655d78c5d4ad05eac287f23a435924df7144.
> >>> This commit introduced a regression that can cause mount hung.
> >>> The changes in __ocfs2_find_empty_slot causes that any node with
> >>> none-zero node number can grab the slot that was already taken by
> >>> node 0, so node 1 will access the same journal with node 0, when it
> >>> try to grab journal cluster lock, it will hung because it was already
> >>> acquired by node 0.
> >>> It's very easy to reproduce this, in one cluster, mount node 0 first,
> >>> then node 1, you will see the following call trace from node 1.
> >>
> >>  From your description, it looks your env mixed local-mount & clustered-mount.
> > No, only cluster mount.
> >>
> >> Could you mind to share your test/reproducible steps.
> >> And which ha stack do you use, pmck or o2cb?
> >>
> >> I failed to reproduce it, my test steps (with pcmk stack):
> >> ```
> >> node1:
> >> mount -t ocfs2 /dev/vdd /mnt
> >>
> >> node2:
> >> for i in {1..100}; do
> >> echo "mount <$i>"; mount -t ocfs2 /dev/vdd /mnt;
> >> sleep 3;
> >> echo "umount"; umount /mnt;
> >> done
> >> ```
> >>
> > Try set one node with node number 0 and mount it there first. I used o2cb stack.
> 
> Could you show more test info/steps. I can't follow your meaning.
> How to set up a node with a fix node number?
> With my understanding, under pcmk env, the first mounted node will auto got node
> number 1 (or any value great than 0). and there is no place to set node number
> by hand. It's very likely you mixed to use nocluster & cluster mount.
> If my suspect right (mixed mount), your use case is wrong.
> 
> >> This local mount feature helps SUSE customers to maintain ocfs2 partition, it's useful.
> >> I want to find whether there is a idear way to fix the hung issue.
> >>
> >>> [13148.735424] INFO: task mount.ocfs2:53045 blocked for more than 122 seconds.
> >>> [13148.739691]       Not tainted 5.15.0-2148.0.4.el8uek.mountracev2.x86_64 #2
> >>> [13148.742560] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> >>> [13148.745846] task:mount.ocfs2     state:D stack:    0 pid:53045 ppid: 53044 flags:0x00004000
> >>> [13148.749354] Call Trace:
> >>> ...
> >>> To fix it, we can just fix __ocfs2_find_empty_slot. But original commit
> >>> introduced the feature to mount ocfs2 locally even it is cluster based,
> >>> that is a very dangerous, it can easily cause serious data corruption,
> >>> there is no way to stop other nodes mounting the fs and corrupting it.
> >>
> >> I can't follow your meaning. When users want to use local mount feature, they MUST know
> >> what they are doing, and how to use it.
> > I can’t agree with you. There is no  mechanism to make sure customer will follow that, you can’t expect customer understand tech well or even read the doc.
> 
> yes, no one reads doc by default.
> 
> currently, mount with option 'nocluster' will show special info to user:
> 
> ```
> # mount -t ocfs2 -o nocluster /dev/vdd /mnt
> Warning: to mount a clustered volume without the cluster stack.
> Please make sure you only mount the file system from one node.
> Otherwise, the file system may be damaged.
> Proceed (y/N):
> ```
> 
> > It’s not the case that you don’t have choice, setup cluster stack is the way to stop customer doing something bad, I believe you have to educate customer to understand this is the cost to guard data security, otherwise when something bad happens, they will lose important data, maybe even no way to recover.
> 
> This feature is not enabled by default, and also shows enough info/warn before executing.
> I give (may awkward) another example:
> nocluster mount likes executing command 'rm -rf /', do you think we should
> tell/educate customer do not execute it?
> 
> The nocluster mount feature was designed to resolve customer pain point from real world:
> SUSE HA stack uses pacemaker+corosync+fsdlm+ocfs2, which complicates/inconveniences
> to set up. and need to install dozens of related packages.
> 
> The nocluster feature main use case:
> customer wants to avoid to set up HA stack, but they wants to check ocfs2 volume
> or do backup volume.
> 
> In my opinion, we should make ocfs2 more powerful and include more useful features for users.
> If there are some problems related new feature, we should do our best to fix it not revert it.

I am not familiar with o2cb stack. If o2cb could give a node with node
number ZERO, I have an idea to avoid mixed noncluster & cluster mounting.

there is slot map management struct:

struct ocfs2_extended_slot {
/*00*/	__u8	es_valid;
	__u8	es_reserved1[3];
	__le32	es_node_num;
/*08*/
};

we could use the es_reserved1[0] to give nocluster mount a speical flag.
maybe we could define:

#define OCFS2_NOCLUSTER_MOUNT 1
if (XX->es_reserved1[0] == OCFS2_NOCLUSTER_MOUNT)
	this_slot_is_mounted_by_noclustered_mode;

the code logic:
- When nocluster mount, check the es_valid for existing clustered mount,
  check the es_reserved1[0] for existing noclustered mount.
  If no other nodes mounted, do the noclustered mount, and set es_reserved1[0]
  with OCFS2_NOCLUSTER_MOUNT. Then clear this value when unmount.
- When another node prepares to mount with clustered mode, it should check
  es_reserved1[0] for detecting noclustered mount. ocfs2 should block the mount
  action if any slot is marked with OCFS2_NOCLUSTER_MOUNT. (make noclustered
  mount unique)

Thanks,
Heming

> 
> >>
> >>  From mount.ocfs2 (8), there also writes *only* mount fs on *one* node at the same time.
> >> And also tell user fs will be damaged under wrong action.
> >>
> >> ```
> >> nocluster
> >>
> >>   This  option  allows  users  to  mount a clustered volume without configuring the cluster
> >>   stack.  However, you must be aware that you can only mount the file system from one  node
> >>   at the same time, otherwise, the file system may be damaged. Please use it with caution.
> >> ```
> >>
> >>> Setup ha or other cluster-aware stack is just the cost that we have to
> >>> take for avoiding corruption, otherwise we have to do it in kernel.
> >>
> >> It's a little bit serious to totally revert this commit just under lacking sanity
> >> check. If you or maintainer think the local mount should do more jobs to prevent mix
> >> local-mount and clustered-mount scenario, we could add more sanity check during
> >> local mounting.
> > I don’t think this should be done in kernel. Setup cluster stack is the way to forward.
> > 
> 
> my mistake: all above 'local mount' should be 'nocluster mount'.
> 
> At last, let's totally understand your use case (or reproduce your hung issue).
> 
> Thanks,
> Heming
>