Re: [PATCH] PM / hibernate: Introduce snapshot test mode for hibernation

[Rafael J. Wysocki]

On Wed, Jul 13, 2016 at 11:45 PM, Pavel Machek <pavel@xxxxxx> wrote:
> On Wed 2016-07-13 22:44:24, Rafael J. Wysocki wrote:
>> On Wed, Jul 13, 2016 at 10:26 PM, Pavel Machek <pavel@xxxxxx> wrote:
>> > On Wed 2016-07-13 22:04:27, Rafael J. Wysocki wrote:
>> >> On Wed, Jul 13, 2016 at 7:01 PM, Pavel Machek <pavel@xxxxxx> wrote:
>> >> > Hi!
>> >> >
>> >> >> >>and then swapon the swap device, and do a testing. This should be safer?
>> >> >> >Yeah, that's the way. Read-only root is other option.
>> >> >> >
>> >> >> >>>I guess updating documentation would be welcome from my side,
>> >> >> >>>otherwise it should be ok.
>> >> >> >>OK, I'll  update the documents.
>> >> >> >Just add fat warning into the documentation.
>> >> >> OK.
>> >> >
>> >> > Actually... If you could add
>> >> >
>> >> > printk(KERN_ALERT "Hibernation image written. If you have any
>> >> > filesystems mounted read-write and attempt to resume, you'll corrupt
>> >> > your data. To prevent that, remove the hibernation image.\n")
>> >> >
>> >> > ...I guess that would save someone's filesystem. (Yes, very high
>> >> > loglevel. If you attempt to do this from anything else then singleuser
>> >> > or initrd, you are asking for problems, so... lets make sure user sees
>> >> > it.)
>> >>
>> >> Please see the new version of this patch:
>> >> https://patchwork.kernel.org/patch/9226837/
>> >
>> > New version changes nothing, right? You still need to be sure
>> > filesystems are not mounted r/w. So I would still like to see printk()
>> > with warning.
>>
>> It shouldn't matter how they are mounted, because the contents of
>> persistent storage don't change.
>
> @@ -721,6 +724,9 @@  int hibernate(void)
>    atomic_inc(&snapshot_device_available);
>      Unlock:
>         unlock_system_sleep();
>         +       if (snapshot_test)
>         +          software_resume();
>         +
>                 return error;
>                  }
>
> Aha, I see, immediate wakeup here. Makes sense. ... ...
>
> No.
>
> AFAICT, freezer is used in hibernation_snapshot, which means at
> Unlock:, kernel threads are running; software_resume() freezes them
> again, but they had chance to run and potentially corrupt the
> persistent storage... right?

OK, there is a bug.

The thawing of user space is potentially dangerous, so in the
"snapshot" test mode hibernate() should just call
free_basic_memory_bitmaps() and from there invoke the code below the
Check_image label in software_resume(), roughly.

Thanks,
Rafael