[pbs-devel] [PATCH FOLLOW-UP proxmox-backup 2/4] task tracking: actually reset entry if desynced

[Hannes Laimer]

On 11/20/25 11:22, Fabian Grünbichler wrote:
> On November 20, 2025 10:37 am, Hannes Laimer wrote:
>> hmm, I'm not sure pushing a new 0/0 entry in that case adds much...
>> logging this though makes a lot if sense
>>
>> actually, I think my patch is not correct. If we have `0/0` and call
>> update with -1 we'd end up with a -1 count in the tracking file.
>> decrementing is also a problem with a 0 counter, not just with
>> non-existing entries.
> 
> that's true. maybe we should first answer the question how we want to
> handle such a mismatch, and then think about implementation details ;)
> 
> AFAICT:
> 
> - we add an operation during datastore lookup (two calls)
> - we add an operation when cloning a datastore instance (one call)
> - we remove an operation when dropping a datastore instance (one call)
> 
> there's some more which are only used by examples and should maybe be
> dropped..
> 
> if a process crashes without executing the drop handler, a left-over
> entry could exist. but such an entry will be cleaned up by the next
> update_active_operations call since the PID is no longer valid.
> 
> so the only remaining issues would be:
> - explicitly leaking instead of dropping a datastore (should never be
>    done)
> - manually editing the active operations file
> - unlinking the lock file while it is used
> 
> effectively, if we would ever end up with an active operation count < 0
> for a given PID, we know something is wrong. but we can not recover for
> this particular PID, so maybe we should add a poison flag (or use a
> negative count as such), and require that process to exit before
> considering the datastore to be "sane" again?
> 

yes, treating a negative value as such makes sense. I guess in such a
case we just shouldn't allow the creating of new (IO) tasks completely.
This should allow running tasks to finish, and with a log message saying
something like
`looks like something went wrong... please restart proxy.service`

not sure if lookup's are fine, since we don't really enforce that the
reference is actually never used for IO...
([1] from some time ago would address that, but only orthogonality
relevant here)


[1] 
https://lore.proxmox.com/pbs-devel/20250526141445.228717-1-h.laimer@proxmox.com/

> there are only a few places where the operation counts matter:
> - removal from the cache map to close FDs when the last task exits, in
>    case certain maintenance mode is set
> - waiting for active tasks to be done before activating certain
>    maintenance modes
> 
> neither of this can be done (safely) if we can no longer tell whether
> there are active tasks..
> 
>>
>> On 11/20/25 10:03, Fabian Grünbichler wrote:
>>> and warn about it. this *should* never happen unless the tracking file got
>>> somehow messed with manually..
>>>
>>> Signed-off-by: Fabian Grünbichler <f.gruenbichler at proxmox.com>
>>> ---
>>> This one fixes the replied-to patch to also correctly store an entry with no
>>> tasks for the current PID, instead of just returning that there are none..
>>>
>>> I am actually not sure how we should handle such a desync, we now pretend it's
>>> the last task even though we don't know for sure.. maybe we should just error
>>> out and let the Drop handler (not) handle it?
>>>
>>>    pbs-datastore/src/task_tracking.rs | 12 ++++++++++--
>>>    1 file changed, 10 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/pbs-datastore/src/task_tracking.rs b/pbs-datastore/src/task_tracking.rs
>>> index 10afebbe2..755d88fdf 100644
>>> --- a/pbs-datastore/src/task_tracking.rs
>>> +++ b/pbs-datastore/src/task_tracking.rs
>>> @@ -94,7 +94,7 @@ pub fn get_active_operations_locked(
>>>    pub fn update_active_operations(
>>>        name: &str,
>>>        operation: Operation,
>>> -    count: i64,
>>> +    mut count: i64,
>>>    ) -> Result<ActiveOperationStats, Error> {
>>>        let path = PathBuf::from(format!("{}/{}", crate::ACTIVE_OPERATIONS_DIR, name));
>>>    
>>> @@ -131,7 +131,15 @@ pub fn update_active_operations(
>>>            None => Vec::new(),
>>>        };
>>>    
>>> -    if !found_entry && count > 0 {
>>> +    if !found_entry {
>>> +        if count < 0 {
>>> +            // if we don't have any operations at the moment, decrementing is not possible..
>>> +            log::warn!(
>>> +                "Active operations tracking mismatch - no current entry for {pid} but asked
>>> +to decrement by {count}!"
>>> +            );
>>> +            count = 0;
>>> +        };
>>>            match operation {
>>>                Operation::Read => updated_active_operations.read = count,
>>>                Operation::Write => updated_active_operations.write = count,
>>
>>