Re: [PATCH] mm: fix account pmd page to the process

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Thu, Jun 16, 2016 at 09:47:46AM -0700, Mike Kravetz wrote:
> On 06/16/2016 09:31 AM, Michal Hocko wrote:
> > On Thu 16-06-16 09:05:23, Mike Kravetz wrote:
> >> On 06/16/2016 08:43 AM, Michal Hocko wrote:
> >>> [It seems that this patch has been sent several times and this
> >>> particular copy didn't add Kirill who has added this code CC him now]
> >>>
> >>> On Thu 16-06-16 17:42:14, Michal Hocko wrote:
> >>>> On Thu 16-06-16 19:36:11, zhongjiang wrote:
> >>>>> From: zhong jiang <zhongjiang@huawei.com>
> >>>>>
> >>>>> when a process acquire a pmd table shared by other process, we
> >>>>> increase the account to current process. otherwise, a race result
> >>>>> in other tasks have set the pud entry. so it no need to increase it.
> >>>>>
> >>>>> Signed-off-by: zhong jiang <zhongjiang@huawei.com>
> >>>>> ---
> >>>>>  mm/hugetlb.c | 5 ++---
> >>>>>  1 file changed, 2 insertions(+), 3 deletions(-)
> >>>>>
> >>>>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> >>>>> index 19d0d08..3b025c5 100644
> >>>>> --- a/mm/hugetlb.c
> >>>>> +++ b/mm/hugetlb.c
> >>>>> @@ -4189,10 +4189,9 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
> >>>>>  	if (pud_none(*pud)) {
> >>>>>  		pud_populate(mm, pud,
> >>>>>  				(pmd_t *)((unsigned long)spte & PAGE_MASK));
> >>>>> -	} else {
> >>>>> +	} else 
> >>>>>  		put_page(virt_to_page(spte));
> >>>>> -		mm_inc_nr_pmds(mm);
> >>>>> -	}
> >>>>
> >>>> The code is quite puzzling but is this correct? Shouldn't we rather do
> >>>> mm_dec_nr_pmds(mm) in that path to undo the previous inc?
> >>
> >> I agree that the code is quite puzzling. :(
> >>
> >> However, if this were an issue I would have expected to see some reports.
> >> Oracle DB makes use of this feature (shared page tables) and if the pmd
> >> count is wrong we would catch it in check_mm() at exit time.
> >>
> >> Upon closer examination, I believe the code in question is never executed.
> >> Note the callers of huge_pmd_share.  The calling code looks like:
> >>
> >>                         if (want_pmd_share() && pud_none(*pud))
> >>                                 pte = huge_pmd_share(mm, addr, pud);
> >>                         else
> >>                                 pte = (pte_t *)pmd_alloc(mm, pud, addr);
> >>
> >> Therefore, we do not call huge_pmd_share unless pud_none(*pud).  The
> >> code in question is only executed when !pud_none(*pud).
> > 
> > My understanding is that the check is needed after we retake page lock
> > because we might have raced with other thread. But it's been quite some
> > time since I've looked at hugetlb locking and page table sharing code.
> 
> That is correct, we could have raced. Duh!
> 
> In the case of a race, the other thread would have incremented the
> PMD count already.  Your suggestion of decrementing pmd count in
> this case seems to be the correct approach.  But, I need to think
> about this some more.

Yes, I made mistake by increasing nr_pmds, not descreasing here.

Testcase:

#include <errno.h>
#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <unistd.h>
#include <sys/mman.h>
#include <sys/syscall.h>
#include <sys/time.h>

#define HPGSZ 2097152UL
int main(int argc, char **argv) {
	char *addr;

	system("echo 1024 > /proc/sys/vm/nr_hugepages");
	addr = mmap(NULL, 1024*HPGSZ, PROT_WRITE | PROT_READ,
			MAP_SHARED | MAP_ANONYMOUS | MAP_HUGETLB | MAP_POPULATE, -1, 0);
	if (addr == MAP_FAILED) {
		fprintf(stderr, "Failed to alloc hugepage\n");
		return -1;
	}

	addr[0] = 1;
	fork();
	printf("addr[0]: %d\n", addr[0]);

	sleep(1);
	return 0;
}

You can simulate race by replacing 'if (pud_none(*pud))' with "if (0)". It
would produce "BUG: non-zero nr_pmds on freeing mm: 2" on the test-case.

Fix:

>From fd22922e7b4664e83653a84331f0a95b985bff0c Mon Sep 17 00:00:00 2001
From: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Date: Fri, 17 Jun 2016 15:07:03 +0300
Subject: [PATCH] hugetlb: fix nr_pmds accounting with shared page tables

We account HugeTLB's shared page table to all processes who share it.
The accounting happens during huge_pmd_share().

If somebody populates pud entry under us, we should decrease pagetable's
refcount and decrease nr_pmds of the process.

By mistake, I increase nr_pmds again in this case. :-/
It will lead to "BUG: non-zero nr_pmds on freeing mm: 2" on process'
exit.

Let's fix this by increasing nr_pmds only when we're sure that the page
table will be used.

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: zhongjiang <zhongjiang@huawei.com>
Fixes: dc6c9a35b66b ("mm: account pmd page tables to the process")
Cc: <stable@vger.kernel.org>        [4.0+]
---
 mm/hugetlb.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index e197cd7080e6..ed6a537f0878 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -4216,7 +4216,6 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 		if (saddr) {
 			spte = huge_pte_offset(svma->vm_mm, saddr);
 			if (spte) {
-				mm_inc_nr_pmds(mm);
 				get_page(virt_to_page(spte));
 				break;
 			}
@@ -4231,9 +4230,9 @@ pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
 	if (pud_none(*pud)) {
 		pud_populate(mm, pud,
 				(pmd_t *)((unsigned long)spte & PAGE_MASK));
+		mm_inc_nr_pmds(mm);
 	} else {
 		put_page(virt_to_page(spte));
-		mm_inc_nr_pmds(mm);
 	}
 	spin_unlock(ptl);
 out:
-- 
 Kirill A. Shutemov