16.10.2012 13:03, Eric W. Biederman пишет:
Stanislav Kinsbursky <[email protected]> writes:
15.10.2012 23:47, Eric W. Biederman пишет:
Hmm. Come to think of it I don't see why you need to set the id at all.
We are using an idr allocator which effectively offers the semantics
that the lowest available id will be allocated. The same semantics we
have for file descriptors.
So it should be possible at least for the first pass at
checkpoint/restart to implement the restoration of sysv ipc without
IPC_PRESET at all.
So IPC_PRESET should just be an optimization, not a necessary feature.
CRIU was designed to suspend/restore not only containers with it's own IPC
namespace, but also for single process and process tree.
So we have to restore IPC objects with proper id and key.
I was not suggesting restoring IPC objects without the proper id and
key.
I was pointing that since the algorithm for id assignment is known, that
by simply allocating objects with care you can guarantee that you
allocate objects with the proper id and key.
Does it means, that you suggest to allocate as many objects, as required to get
desired id and then release all, that not needed?
As for the case when you don't have your own ipc namespace, in general
that case will fail. Especially with the current id assignment
algorithm chances that your id is already taken are very high. So since
in general and most of the time it will fail to restore an ipc namespace
that is not shared outside of the container there needs to be a
compelling reason to support to even consider it.
The reason is simple: we want at least to have a chance to restore such
processes with IPC objects.
Yes, in general, resource id will be close to 0, and it would be most probably
impossible to restore such resource on another machine, if any software on it
uses IPC objects.
But this doesn't mean, that the implementation is useless.
Imagine, that you have a process tree, sharing some IPC resource. And it's
actually all additional software, that you have running on a hardware node
(except processes, started during boot by default).
Then you have another node with the same configuration. And you want to use them
for load balancing or something like this. I.e. you want to migrate this process
tree from one to another.
In this particular case (which is valid and used widely) this migration will be
always successful.
I am not comfortable with the key<->id association changing on active
objects. That gives rise to new and scary races in userspace, and a
much more complex set of interactions to analyze to see if the both
userspace and the kernel code is correct.
Could you, please, give an example, what scares you?
The case, when some software can change IPC object key, thus making some other
task unable to attach this object by key, is not significant, because task can
destroy object by key already.
I am not particularly comfortable adding a new allocation pathway for
sysv ipc objects when that new pathway appears completely unnecessary.
Stanislav please concentrate on something that works for the easy case
of an ipc namespace per container today. Once there is a working
checkpoint/restart and people can use it and see what the limitations
it becomes much easier to see if extra optimizations like your
IPC_PRESET code are worth it.
Sorry, but I don't understand your objection.
Eric
--
Best regards,
Stanislav Kinsbursky
_______________________________________________
Devel mailing list
[email protected]
https://openvz.org/mailman/listinfo/devel
