lkml.org 
[lkml]   [2009]   [May]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 18/38] C/R: core stuff


Alexey Dobriyan wrote:
> On Wed, May 27, 2009 at 06:45:04PM -0400, Oren Laadan wrote:
>> Alexey Dobriyan wrote:
>>> On Wed, May 27, 2009 at 04:56:27PM -0400, Oren Laadan wrote:
>>>> Alexey Dobriyan wrote:
>>>>> On Tue, May 26, 2009 at 08:16:44AM -0500, Serge E. Hallyn wrote:
>>>>>> Quoting Alexey Dobriyan (adobriyan@gmail.com):
>>>>>>> Introduction
>>>>>>> ------------
>>>>>>> Checkpoint/restart (C/R from now) allows to dump group of processes to disk
>>>>>>> for various reasons like saving process state in case of box failure or
>>>>>>> restoration of group of processes on another or same machine later.
>>>>>>>
>>>>>>> Unlike, let's say, hypervisor C/R style which only needs to freeze guest kernel
>>>>>>> and dump more or less raw pages, proposed C/R doesn't require hypervisor.
>>>>>>> For that C/R code needs to know about all little and big intimate kernel details.
>>>>>>>
>>>>>>> The good thing is that not all details needs to be serialized and saved
>>>>>>> like, say, readahead state. The bad things is still quite a few things
>>>>>>> need to be.
>>>>>> Hi Alexey,
>>>>>>
>>>>>> the last time you posted this, I went through and tried to discern the
>>>>>> meaningful differences between yours and Oren's patchsets. Then I sent some
>>>>>> patches to Oren to make his set configurable to act more like yours. And Oren
>>>>>> took them! But now you resend this patchset with no real changelog, no
>>>>>> acknowledgment that Oren's set even exists
>>>>> Is this a requirement? Everybody following topic already knows about
>>>>> Oren's patchset.
>>>> Some people do ack other people's work. See for example patches #1
>>>> and #24 in my recent post. You're welcome.
>>>>
>>>>>> - or is much farther along and pretty widely reviewed and tested (which is
>>>>>> only because he started earlier and, when we asked for your counterpatches
>>>>>> at an earlier stage, you would never reply) - or, most importantly, what
>>>>>> it is that you think your patchset does that his does not and cannot.
>>>>> There are differences. And they're not small like you're trying to describe
>>>>> but pretty big compared the scale of the problem.
>>>> I've asked before, and I repeat now: can you enumerate these "big"
>>>> scary differences that make it such a "big" problem ?
>>>>
>>>> So far, we identified two main "design" issues -
>>> Why in "? Yes, they are high-level design issues.
>>>
>> In quotes, because I argued further on that, although my patchset
>> takes a stand on both issues, it can be easily reverted _within_
>> that patchset. Moreover, I argue that they can co-exist.
>>
>>>> 1) Whether or not allow c/r of sub-container (partial hierarchy)
>>>>
>>>> 2) Creation of restarting process hierarchy in kernel or in userspace
>>>>
>>>> As for #1, you are the _only_ one who advocates restricting c/r to
>>>> a full container only. I guess you have your reasons, but I'm unsure
>>>> what they may be.
>>> The reason is that checkpointing half-frozen, half-live container is
>>> essentially equivalent to live container which adds much complexity
>>> to code fundamentally preventing kernel from taking coherent snapshot.
>>>
>>> In such situations kernel will do its job badly.
>> In such situation the kernel will do a bad job if the user is asking
>> for a bad job.
>
> User doesn't even understand why we're discussing this issue so hard.
>
>> Just like checkpointing without snapshotting the file system and expecting
>> it to always work.
>
> This is different.
>
> Kernel can't do anything about not-synced fs. Because nodoby is
> advocating that kernel should sync fs. Consequently, screwup in fs sync is
> clearly user failure. Any (yours, mine) in-kernel C/R has this failure mode,
> so we skip it and discuss what's left.
>
> Now, kernel CAN do something about tasks and other data structures
> because it easily controls them.
>
> Your procedure for checkpointing starts with "kill -STOP".

Wrong. It requires the processes to be frozen.

> To make anything reliable, you have to ban "kill -CONT" for the duration of
> checkpointing. Is this done BTW? I don't remember new flags added
> in task_struct. Or this is going to be skipped on grounds that it's
> user screwup (potentially oopsable).
>
> That's why, OpenVZ relies on suspend-to-ram freezer solely, because userspace
> can't arbitrarily send suspend and freeze notifications. We only need to
> protect against untimely STR unfreeze which only adds code in C/R code
> not in task_struct.

Same principle for both patchsets: tasks may *not* be permitted to
execute while being checkpointed.

For this I suggested a CHECKPOINTING freezer state: transition to/from
this state is done _only_ by sys_checkpoint(), so that checkpointed
processes cannot be unfrozen. Matt Helseley already posted a patch to
implement this.

>
> But you aren't going to stop with task_struct, and add such flags all
> over the place. Or you're going to declare that it's all user fault.

I'm unsure what sort of flags you have in mind, or what's the problem
that alleged flags would solve ? So far I didn't plan any.

>
>> But if the user is a bit more careful (and even then, not that much),
>> she can enjoy the wonderful benefits of c/r without the wonderful
>> benefits of containers.
>>
>> If useful, it's easy to pass a flag to checkpoint() that will ask
>> to enforce, say, shared memory "leaks" but not nsproxy or file "leaks".
>>
>> In fact, even shared memory "leaks" may be useful for some users (e.g.
>> what the guys from kerlabs pointed out).
>>
>>> Manpage will be filled with strings like "if $FOO is shared then $BAR is
>>> not guaranteed".
>>>
>>> What to do if user simply doesn't know if container is bounded?
>>> Checkpoint and to hell with consequences?
>>>
>>> If two tasks share mm_struct you can't even detect that pages you dump
>>> aren't filled with garbage meanwhile from second task.
>>>
>>> If two tasks share mm_struct, other task can issue AIO indefinitely
>>> preventing from taking even coherent filesystem snapshot.
>>>
>>> That's why I raise this issue again to hear from people what they think
>>> and these people shouldn't be containers and C/R people, because the
>>> latter already made up their minds.
>> Lol .. and disagreement persists among us :)
>>
>> And indeed, I have heard and seen already a few opinions in favor
>> of permitting non-container checkpoint. From potential users (not
>> c/r people).
>>
>>> This is super-important issue to get right from the beginning.
>>>
>>>> On the other hand, there has been a handful of use-cases and opinions
>>>> in favor of allowing both capabilities to co-exist. Not the mention
>>>> that nearly no additional code is necessary, on the contrary.
>>>>
>>>> As for #2, you didn't even bother to reply to the discussion that I
>>>> had started about it. This decision is important to allow future
>>>> flexibility of the mechanism, and to address the needs of several
>>>> potential users, as seen in that discussion and others. Here, too,
>>>> you are the _only_ one that advocates that direction.
>>> Are you going to fork to-become-zombies, make them call restart(2) and
>>> zombify?
>> Yes.
>>
>>>> And the funniest thing -- *both* decisions can be *easily* overturned
>>>> in my patchset. In fact, regarding #2 - either way can be easily done
>>>> in it.
>>>>
>>>> So I wonder, what are the "big" issues that bother you so much ?
>>>> "if there is a will, there is a way".
>>> Oren, don't you really understand?
>>>
>>> Users want millions of things, but every thing has price.
>> I beg to differ: there is marginal price to support both --
>
> Wrong, try to do netns with sockets without deadlocks.

Let's indeed discuss netns - that's an important issue. But it's
also huge - "do netns" may mean different things to different
people. How about you start a new thread explaining why it will
be such a problem ?

>
> Price in loose case is strictly bigger than price in strict case
> because set of loose states is strict superset of set of strict cases.

The code to enforce full-container approach is a *superset* of the
code without this requirement.

If you can checkpoint a full container, surely you can checkpoint a
sub-hierarchy of processes.

We both agree that the sub-hierarchy case is not idiot-proof: it's
garbage in, garbage out. In the worst case, the restart will simply
fail, but no other harm done.

(And if you are concerned about security, well - that's a different
topic - and one can disable c/r for non privileged users).

>
>> in fact, enforcing the container requirement (e.g. leaks detection -
>> which, btw, is imperfect and cannot be made race-free)
>
> It'll only generate false positives when say /proc/$PID/maps is being
> read from outside, so mm->mm_users is elevated. I can return -EAGAIN.
>
> If file is opened via /proc/*/fd, proc_fd_access_allowed() is only
> needed to be tweaked and use task_struct field, not new struct file
> field.
>
> ->mmap_sem is taken only for reading in /proc, so fine.
>
> /proc/*/mem -- tweak mem_open().
>
> struct pid, itself -- refcount can be elevated because someone chdir'ed
> into /proc/*, but I don't do refcount checks on struct pid, because it's
> hard (inode in memory pins struct pid) and pids numbers themselves don't
> change once taken. So, if struct pid is somehow leaked to outside,
> there will be NO incoherency in image and no problems with locking or
> whatever.
>
> For netns elevated refcount, get_proc_task_net() will also reject
> open(2) request based on task_struct field, not netns field!
>
> I'm sorry, but all bases re refcount checks are pretty well covered
> and are easily closed with tiny bit of code.
>
> Please show counter-examples where after all data structures are
> collected and tasks being frozen and refcount checks passed,
> userspace can still access one of them for writing.

ptrace ? (sure, can also be "tweaked" to disallow access while
task a checkpointing. Everything can :)

>
>> *adds* code over the non-container case.
>
> Whole 8 lines per data-structure:
>
> for_each_kstate_object(ctx, obj, KSTATE_CTX_MM_STRUCT) {
> struct mm_struct *mm = obj->o_obj;
> unsigned int cnt = atomic_read(&mm->mm_users);
>
> if (obj->o_count + 1 != cnt) {
> pr_err("mm_struct %p has external references %lu:%u\n", mm, obj->o_count, cnt);
> return -EINVAL;
> }
> }
>
> No locking, no nothing, just counting refcounts.
>
>> So in a sense, we get the no-container case for free.
>
> Not for free definitely.
>

s/Not// :)
If you take out (or simply skip) the isolation/coherency tests for
a full-container, you are left with ... sub-hierarchy support. Voila.

Oren.



\
 
 \ /
  Last update: 2009-05-29 00:25    [from the cache]
©2003-2011 Jasper Spaans