Re: [RFC v17][PATCH 22/60] c/r: external checkpoint of a task other than ourself

[Oren Laadan <orenl@librato.com>]

Serge E. Hallyn wrote:
> Quoting Oren Laadan (orenl@librato.com):
>>
>> Serge E. Hallyn wrote:
>>> Quoting Oren Laadan (orenl@librato.com):
>>>> Now we can do "external" checkpoint, i.e. act on another task.
>>> ...
>>>
>>>>  long do_checkpoint(struct ckpt_ctx *ctx, pid_t pid)
>>>>  {
>>>>  	long ret;
>>>>
>>>> +	ret = init_checkpoint_ctx(ctx, pid);
>>>> +	if (ret < 0)
>>>> +		return ret;
>>>> +
>>>> +	if (ctx->root_freezer) {
>>>> +		ret = cgroup_freezer_begin_checkpoint(ctx->root_freezer);
>>>> +		if (ret < 0)
>>>> +			return ret;
>>>> +	}
>>> Self-checkpoint of a task in root freezer is now denied, though.
>>>
>>> Was that intentional?
>> Yes.
>>
>> "root freezer" is an arbitrary task in the checkpoint subtree or
>> container. It is used to verify that all checkpointed tasks - except
>> for current, if doing self-checkpoint - belong to the same freezer
>> group.
>>
>> Since current is busy calling checkpoint(2), and since we only permit
>> checkpoint of (cgroup-) frozen tasks, then - by definition - it cannot
>> possibly belong to the same group. If it did, it would itself be frozen
>> like its fellows and unable to call checkpoint(2).
> 
> So then you're saying that regular self-checkpoint no longer works,
> but the documentation still shows self.c and claims it should just
> work.

I'm unsure why you say that self-checkpoint no longer works ?
In fact, I just double checked that it does.

Self-checkpoint has two immediate use-cases:

1) Single process that checkpoints itself - ctx->root_freezer remains
NULL, which causes cgroup_freezer_begin_checkpoint() to be skipped.

2) Process P that belongs to a hierarchy (subtree or container), and
P calls checkpoint(2) to checkpoint the hierarchy.
For this to work, all other processes in the hierarchy must be frozen.
Therefore, they also belong to a freezer cgroup (perhaps more than one -
but that is not permitted).
In this case, ctx->root will point to a process from the freezer cgroup,
and the code tests all other processes (excluding P, which is current)
to confirm that they belong to the same freezer cgroup.
P itself can not possibly belong to it, otherwise it would have been
frozen and not executing the checkpoint(2) syscall.

IOW, for case 2 to work, one must arrange for all tasks in the target
hierarchy, except for P (- current, the checkpointer), to belong to
a single freezer cgroup, and for that cgroup to be frozen.

>>> Self-checkpoint of a task in root freezer is now denied, though.

Maybe I didn't really understand what you meant by that, and by
"root freezer" ?

> 
> Mind you I prefer this as it is more consistent, but I thought it
> was something you wanted to support.

Self-checkpoint simply allows a process to checkpoint itself (and
perhaps additional processes too). I never quite understood why you
view it as a source of inconsistency ...

Nevertheless, it still works.

Oren.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>