Serge E. Hallyn wrote:
> Quoting Dan Smith (danms@xxxxxxxxxx):
>> This patch adds basic checkpoint/restart support for AF_UNIX sockets. It
>> has been tested with a single and multiple processes, and with data inflight
>> at the time of checkpoint. It supports both socketpair()s and path-based
>> sockets.
>>
>> I have an almost-working AF_INET follow-on to this which I can submit after
>> this is reviewed and tweaked into acceptance.
>>
>> Signed-off-by: Dan Smith <danms@xxxxxxxxxx>
[...]
>
>> +static int sock_un_checkpoint(struct ckpt_ctx *ctx,
>> + struct sock *sock,
>> + struct ckpt_hdr_socket *h)
>> +{
>> + struct unix_sock *sk = unix_sk(sock);
>> + struct unix_sock *pr = unix_sk(sk->peer);
>> + int new;
>> + int ret;
>> +
>> + h->un.this = ckpt_obj_lookup_add(ctx, sk, CKPT_OBJ_SOCK, &new);
>> + if (h->un.this < 0)
>> + goto out;
>> +
>> + if (sk->peer)
>> + h->un.peer = ckpt_obj_lookup_add(ctx, pr, CKPT_OBJ_SOCK, &new);
>> + else
>> + h->un.peer = 0;
>> +
>> + if (h->un.peer < 0) {
>> + ret = h->un.peer;
>> + goto out;
>> + }
>> +
>> + ret = ckpt_write_obj(ctx, (struct ckpt_hdr *) h);
>> + out:
>> + return ret;
>> +}
>
> in the CHECKPOINT_SUBTREE case do we want to try to ensure that sk->peer
> is owned by another checkpointed task?
What exactly would you like to enforce - that it is "in-use"
by a checkpointed task, or that is isn't "in-use" outside ?
It probably makes sense to verify that the socket is "in-use"
by at least one task in the checkpointed set (heh... I expect
kerlab guys to argue against forcing this...), and perhaps
issue a warning ?
(Which is not a bad idea - add a ckpt_write_warning() function
that will write a warning in the image, but won't abort the
entire checkpoint).
It isn't easy to verify the "in-use" property - what if task
A transfers a file using unix-domain sockets to task B (both
in the set), and A closed the file descriptor.... so we can
know it's in transit, but we don't know who will receive the
file eventually.
(Ahh.. of course .. issue a warning :)
It makes less sense to verify the socket is _not_ in use
_outside_ the checkpointed set - and it can be expensive to
do so; After all, there is a whole-container option if you
need that guarantee.
If we are to add such checks, or warnings, it's clearly not
a high priority now (and given akpm's comment ...).
Oren.
_______________________________________________
Containers mailing list
Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linux-foundation.org/mailman/listinfo/containers
