[CRIU] [PATCH v2 3/3] mount: handle tracefs more gracefully

[Pavel Emelyanov]

On 05/13/2016 05:36 PM, Tycho Andersen wrote:
> On Fri, May 13, 2016 at 02:31:09PM +0300, Pavel Emelyanov wrote:
>> On 05/12/2016 09:43 PM, Tycho Andersen wrote:
>>> On Thu, May 12, 2016 at 09:34:10PM +0300, Pavel Emelyanov wrote:
>>>> On 05/12/2016 06:39 PM, Tycho Andersen wrote:
>>>>> On Thu, May 12, 2016 at 05:22:18PM +0300, Pavel Emelyanov wrote:
>>>>>>> +		if (new->fstype->code == FSTYPE__TRACEFS) {
>>>>>>> +			pr_info("\tskipping tracefs mounted at %s\n", new->mountpoint + 1);
>>>>>>> +			mnt_entry_free(new);
>>>>>>> +			new = NULL;
>>>>>>> +			goto end;
>>>>>>> +		}
>>>>>>> +
>>>>>>>  		pr_info("\ttype %s source %s mnt_id %d s_dev %#x %s @ %s flags %#x options %s\n",
>>>>>>>  				fsname, new->source,
>>>>>>>  				new->mnt_id, new->s_dev, new->root, new->mountpoint,
>>>>>>
>>>>>> Why do we need to drop this entry? AFAIU this is only required to make
>>>>>> restore skip mounting one and let 'mount -t debugfs' do the thing, right?
>>>>>> If yes, then we should keep this entry parsed and skip one on restore.
>>>>>> I would say -- the debugfs' -> parse callback should walk all tracefs'
>>>>>> MI's and set mounted on them to true.
>>>>>
>>>>> Hmm, actually I'm a little confused about this. Do you mean ->mount
>>>>> here?
>>>>
>>>> Yup. I wrote ->parse by mistake :(
>>>
>>> The problem here is still that it's a bind mount for us, so debugfs'
>>
>> External bind mount? Is it specified as such with --ext-mount-map?
> 
> With --ext-mount-map auto it gets detected as such.
> 
>>> ->mount never gets called :(
>>
>> How about fixing the tracefs's mi in ->parse callback making it debugfs? :)
> 
> I don't think that would work either, we'd still end up bind mounting
> it and hitting the kernel's overmounting checks.
> 
> What we could do is in debugfs' ->parse to add MS_REC always. We'd
> still need this hunk (although I think it needs to drop the MS_PROPAGATE flags
> in looking at it), but some other extra bits would go away.
> 
> @@ -2474,12 +2488,16 @@ do_bind:
>                 }
>         }
> 
> -       if (mount(root, mi->mountpoint, NULL, MS_BIND, NULL) < 0) {
> +       if (mount(root, mi->mountpoint, NULL, MS_BIND | mi->fstype->flags, NULL) < 0) {
>                 pr_perror("Can't mount at %s", mi->mountpoint);
>                 goto err;
>         }
> 
> 
> but we could get rid of the other hungs and fstype->flags all together. That
> doesn't solve our tracefs problem, but it does make the patch smaller.

Hm.. OK, let's do whatever is required, but try to avoid checks for exact
fstype name/code in generic code.

-- Pavel