Re: [RFC][PATCH 8/8] check files for checkpointability

[Cedric Le Goater <legoater@free.fr>]

Dave Hansen wrote:
> On Mon, 2009-03-02 at 11:44 -0600, Serge E. Hallyn wrote:
>> Quoting Dave Hansen (dave@linux.vnet.ibm.com):
>>> On Mon, 2009-03-02 at 11:22 -0600, Nathan Lynch wrote:
>>>> No.. I mean what if a process 1234 does
>>>>
>>>> f = fopen("/proc/1234/stat", "r");
>>>>
>>>> and is then checkpointed.  Can that path be resolved during restart,
>>>> before pid 1234 is alive?
>>> Heh, that's a good one.
>>>
>>> It does mean that we can't do restore like this:
>>>
>>> 	for_each_cr_task()
>>> 		restore_task_struct()
>>> 		restore_files()
>>> 		...
>>>
>>> We have to do:
>>>
>>> 	for_each_cr_task()
>>> 		restore_task_struct()
>>> 	for_each_cr_task()
>>> 		restore_files()
>>>
>> Which is what we actually do, right?
> 
> OK, I have a really evil one.
> 
> What if task 1234 does:
> 
> 	open(O_RDONLY, "/proc/5678/fdinfo/44");
> 
> and task 5678 does:
> 
> 	open(O_RDONLY, "/proc/5678/fdinfo/55");
> 
> There is no right order.
> 
> The only right way I can think to do it is that we have to loop on the
> restore and defer files that we can't seem to find right now, hoping
> that they'll show up as the restore progresses.

or the restore algorithm should support recursion. for example, epoll, 
attached 'struct files' to af_unix socket, pipes (2 ends), fifos (idem),
connected socket (you need the listening end), etc. 

C.

> Basically:
> 
> 	for_each_cr_task()
> 		deferred_files = restore_files()
> retry:
> 	making_progress = 0
> 	for_each(deferred_file)
> 		restore(deferred_file)
> 	if (making_progress)
> 		goto retry;