Hello.

Thanks for the reply.

We are doing the test (a kind of aging test for 3 days) to prove this is 
the fix for the problem.

I will let you know when the test is done.



On 09/25/2018 06:08 AM, Andrea Arcangeli wrote:

Hello,

On Sat, Sep 22, 2018 at 01:38:07PM +0900, Chulmin Kim wrote:

quoted

Dear Arcangeli,


I think this problem is very much related with

the race condition shown in the below commit.

(e86f15ee64d8, mm: vma_merge: fix vm_page_prot SMP race condition
against rmap_walk)


I checked that

the the thread and its child threads are doing mprotect(PROT_{NONE or
R|W}) things repeatedly

while I didn't reproduce the problem yet.


Do you think this is one of the phenomenon you expected

from the race condition shown in the above commit?

Yes that commit will fix your problem in a v4.4 based tree that misses
that fix. You just need to cherry-pick that commit to fix the problem.

Page migrate sets the pte to PROT_NONE by mistake because it runs
concurrently with the mprotect that transitions an adjacent vma from
PROT_NONE to PROT_READ|WRITE. vma_merge (before the fix) temporarily
shown an erratic PROT_NONE vma prot for the virtual range under page
migration.

With NUMA disabled, it's likely compaction that triggered page migrate
for you. Disabling compaction at build time would have likely hidden
the problem. Compaction uses migration and you most certainly have
CONFIG_COMPACTION=y (rightfully so).

On a side note, I suggest to cherry pick the last upstream commit of
mm/vmacache.c too.

Sorry but I didn't get this line correctly.

Do you meanthe commit 7a9cdebdc (mm: get rid of vmacache_flush_all() 
entirely)?
Could you elaborate what is the point?
Are you saying there is another scenario that makes the problem I am seeing?

Hope this helps,
Andrea

Thanks.
Chulmin Kim