Message-ID: <47A72891.4000404@fr.ibm.com>
Date: Mon, 04 Feb 2008 16:00:33 +0100
From: Daniel Lezcano <dlezcano@fr.ibm.com>
User-Agent: Thunderbird 2.0.0.9 (X11/20071115)
MIME-Version: 1.0
To: Pavel Emelyanov <xemul@openvz.org>
CC: Kirill Korotaev <dev@sw.ru>, containers@lists.linux-foundation.org,
       Cedric Le Goater <clg@fr.ibm.com>, linux-kernel@vger.kernel.org,
       Alexey Dobriyan <adobriyan@gmail.com>
Subject: Re: [Devel] Re: [PATCH 2.6.24-rc8-mm1 09/15] (RFC) IPC: new kernel
 API to change an ID
References: <20080129160229.612172683@bull.net>	<20080129162000.454857358@bull.net>	<20080129210656.GB1990@martell.zuzino.mipt.ru>	<47A18E47.5050206@bull.net>	<47A19AC2.7040709@sw.ru>	<47A1B78C.7050405@bull.net>	<47A1C8FE.9010700@sw.ru> <47A1F2DB.7080600@fr.ibm.com> <47A71606.5030201@sw.ru> <47A71BDF.5000801@openvz.org>
In-Reply-To: <47A71BDF.5000801@openvz.org>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2414
Lines: 53

Pavel Emelyanov wrote:
> Kirill Korotaev wrote:
>> Cedric Le Goater wrote:
>>> Hello Kirill !
>>>
>>> Kirill Korotaev wrote:
>>>> Pierre,
>>>>
>>>> my point is that after you've added interface "set IPCID", you'll need
>>>> more and more for checkpointing:
>>>> - "create/setup conntrack" (otherwise connections get dropped),
>>>> - "set task start time" (needed for Oracle checkpointing BTW),
>>>> - "set some statistics counters (e.g. networking or taskstats)"
>>>> - "restore inotify"
>>>> and so on and so forth.
>>> right. we know that we will have to handle a lot of these
>>> and more and we will need an API for it :) so how should we handle it ?
>>> through a dedicated syscall that would be able to checkpoint and/or
>>> restart a process, an ipc object, an ipc namespace, a full container ?
>>> will it take a fd or a big binary blob ?  
>>> I personally really liked Pavel idea's of filesystem. but we dropped the
>>> thread.
>> Imho having a file system interface means having all its problems.
>> Imagine you have some information about tasks exported with a file system interface.
>> Obviously to collect the information you have to hold some spinlock like tasklist_lock or similar.
>> Obviously, you have to drop the lock between sys_read() syscalls.
>> So interface gets much more complicated - you have to rescan the objects and somehow find the place where
>> you stopped previous read. Or you have to to force reader to read everything at once.
> 
> To remember the place when we stopped previous read we have a "pos" counter
> on the struct file.
> 
> Actually, tar utility, that I propose to perform the most simple migration
> reads the directory contents with 4Kb buffer - that's enough for ~500 tasks.
> 
> Besides, is this a real problem for a frozen container?

I like the idea of a C/R filesystem. Does it implies a specific user 
space program to orchestrate the checkpoint/restart of the different 
subsystems ? I mean the checkpoint is easy but what about the restart ? 
We must ensure, for example to restore a process before restoring the fd 
associated to it, or restore a deleted file before restoring the fd 
opened to it, no ?


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/