On Wed, May 20, 2015 at 1:46 AM, Rafael J. Wysocki <r...@rjwysocki.net> wrote:
> On Monday, May 18, 2015 12:23:00 PM Anisse Astier wrote:
>> Hi Rafael,
>>
>> Thanks for taking the time to review this.
>>
>> On Sat, May 16, 2015 at 2:28 AM, Rafael J. Wysocki <r...@rjwysocki.net>
>> wrote:
>> > On Thursday, May 14, 2015 04:19:46 PM Anisse Astier wrote:
>> >> SANITIZE_FREED_PAGES feature relies on having all pages going through
>> >> the free_pages_prepare path in order to be cleared before being used. In
>> >> the hibernate use case, free pages will automagically appear in the
>> >> system without being cleared, left there by the loading kernel.
>> >>
>> >> This patch will make sure free pages are cleared on resume; when we'll
>> >> enable SANITIZE_FREED_PAGES. We free the pages just after resume because
>> >> we can't do it later: going through any device resume code might
>> >> allocate some memory and invalidate the free pages bitmap.
>> >>
>> >> Signed-off-by: Anisse Astier <ani...@astier.eu>
>> >> ---
>> >> kernel/power/hibernate.c | 4 +++-
>> >> kernel/power/power.h | 2 ++
>> >> kernel/power/snapshot.c | 22 ++++++++++++++++++++++
>> >> 3 files changed, 27 insertions(+), 1 deletion(-)
>> >>
>> >> diff --git a/kernel/power/hibernate.c b/kernel/power/hibernate.c
>> >> index 2329daa..0a73126 100644
>> >> --- a/kernel/power/hibernate.c
>> >> +++ b/kernel/power/hibernate.c
>> >> @@ -305,9 +305,11 @@ static int create_image(int platform_mode)
>> >> error);
>> >> /* Restore control flow magically appears here */
>> >> restore_processor_state();
>> >> - if (!in_suspend)
>> >> + if (!in_suspend) {
>> >> events_check_enabled = false;
>> >>
>> >> + clear_free_pages();
>> >
>> > Again, why don't you do that at the swsusp_free() time?
>>
>> Because it's too late, the kernel has already been through device
>> resume code, and the free pages bitmap isn't valid anymore; device
>> resume code might allocate memory, and we'd be clearing those pages as
>> well.
>
> Are we both talking about the same thing?
I think we aren't talking about the same thing. The free_pages_map is
used for all free pages, plus the memory used for suspend (when it
intersects with forbidden_page_map).
We don't need to clear the memory in swsusp_free, because it already
calls the __free_page code path that clears free pages. What we need,
is to clear pages left by the loader kernel before it jumped into the
resumed kernel.
>
> swsusp_free() is *the* function that, well, frees all the pages allocated
> by the hibernate core, so how isn't the free pages bitmap valid when it is
> called?
Because swsusp_free will only free the intersection of free_pages_map
and forbidden_pages_map. This intersection is the set of pages used
for suspend, and they are in fact allocated, not free. The rest of
free_pages_map are the real free pages, but as I said, once we resume,
it will also include the pages left hanging by the loading kernel.
In addition, the free_pages_map contains a reference to all the free
pages at the moment of suspend, but this reference isn't valid by the
time we reach swsusp_free(), because we've been through many drivers'
resume by then, and those can allocate memory; they might even want
already zeroed memory, which they won't get because of the __GFP_ZERO
optimization in the next patch; we just expect all free pages to be
already cleared, but pages from the loading kernel aren't.
>
> Why don't you add the clearing in there, right at the spot when the pages
> are actually freed?
>
> Moreover, why is the resume code path the only one where freed pages need to
> be sanitized?
Because all pages in the system go through the page freeing path
(which clears them) when leaving the (no)bootmem allocator.
It's a bit long, so I hope that I'm clear in my explanation.
Regards,
Anisse
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
