Re: [PATCH v8 9/9] sparc64: Add support for ADI (Application Data Integrity)

[Khalid Aziz]

On 10/13/2017 08:14 AM, Khalid Aziz wrote:
> On 10/12/2017 02:27 PM, Anthony Yznaga wrote:
> > > On Oct 12, 2017, at 7:44 AM, Khalid Aziz <khalid.aziz@xxxxxxxxxx> wrote:
> > >
> > >
> > > On 10/06/2017 04:12 PM, Anthony Yznaga wrote:
> > > > > On Sep 25, 2017, at 9:49 AM, Khalid Aziz <khalid.aziz@xxxxxxxxxx> 
> > > > > wrote:
> > > > >
> > > > > This patch extends mprotect to enable ADI (TSTATE.mcde), 
> > > > > enable/disable
> > > > > MCD (Memory Corruption Detection) on selected memory ranges, enable
> > > > > TTE.mcd in PTEs, return ADI parameters to userspace and 
> > > > > save/restore ADI
> > > > > version tags on page swap out/in or migration. ADI is not enabled by
> > > > I still don't believe migration is properly supported. Your
> > > > implementation is relying on a fault happening on a page while its
> > > > migration is in progress so that do_swap_page() will be called, but
> > > > I don't see how do_swap_page() will be called if a fault does not
> > > > happen until after the migration has completed.
> > >
> > > User pages are on LRU list and for the mapped pages on LRU list, 
> > > migrate_pages() ultimately calls try_to_unmap_one and makes a 
> > > migration swap entry for the page being migrated. This forces a page 
> > > fault upon access on the destination node and the page is swapped 
> > > back in from swap cache. The fault is forced by the migration swap 
> > > entry, rather than fault being an accidental event. If page fault 
> > > happens on the destination node while migration is in progress, 
> > > do_swap_page() waits until migration is done. Please take a look at 
> > > the code in __unmap_and_move().
> >
> > I looked at the code again, and I now believe ADI tags are never 
> > restored for migrated pages. Here's why:
>
> I will take a look at it again. I have run extensive tests migrating 
> pages of a process across multiple NUMA nodes over and over again and 
> ADI tags were never lost, so this does work. I won't rule out the 
> possibility of having missed a code path where tags are not restored and 
> I will look for it.

Anthony,

I just ran my migration test again which:

- malloc's 16 GB of memory
- Assigns a rotating ADI tag every 64 bytes to the malloc'd buffer
- Writes a pattern to the entire buffer
- Verifies the pattern it wrote using ADI tagged addresses.

While this test was running, I had a script migrate test program pages 
across two NUMA nodes every 30 seconds using migratepages command. I did 
not see an ADI tag mismatch over multiple runs of this test. This test 
shows migration is working.

Can you give me a test that shows the failure you think we should see 
and I will debug it.

Thanks,
Khalid