mm,madvise,hugetlb: fix unexpected data loss with MADV_DONTNEED on hugetlbfs

Message ID 20221021154546.57df96db@imladris.surriel.com
State New
Headers
Series mm,madvise,hugetlb: fix unexpected data loss with MADV_DONTNEED on hugetlbfs |

Commit Message

Rik van Riel Oct. 21, 2022, 7:45 p.m. UTC
  A common use case for hugetlbfs is for the application to create
memory pools backed by huge pages, which then get handed over to
some malloc library (eg. jemalloc) for further management.

That malloc library may be doing MADV_DONTNEED calls on memory
that is no longer needed, expecting those calls to happen on
PAGE_SIZE boundaries.

However, currently the MADV_DONTNEED code rounds up any such
requests to HPAGE_PMD_SIZE boundaries. This leads to undesired
outcomes when jemalloc expects a 4kB MADV_DONTNEED, but 2MB of
memory get zeroed out, instead.

Use of pre-built shared libraries means that user code does not
always know the page size of every memory arena in use.

Avoid unexpected data loss with MADV_DONTNEED by rounding up
only to PAGE_SIZE (in do_madvise), and rounding down to huge
page granularity.

That way programs will only get as much memory zeroed out as
they requested.

While we're here, refactor madvise_dontneed_free_valid_vma
a little so mlocked hugetlb VMAs need MADV_DONTNEED_LOCKED.

Cc: Mike Kravetz <mike.kravetz@oracle.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: David Hildenbrand <david@redhat.com>
Fixes: 90e7e7f5ef3f ("mm: enable MADV_DONTNEED for hugetlb mappings")
---
 mm/madvise.c | 26 +++++++++++++++++++-------
 1 file changed, 19 insertions(+), 7 deletions(-)
  

Comments

Mike Kravetz Oct. 21, 2022, 8:48 p.m. UTC | #1
On 10/21/22 15:45, Rik van Riel wrote:
> A common use case for hugetlbfs is for the application to create
> memory pools backed by huge pages, which then get handed over to
> some malloc library (eg. jemalloc) for further management.
> 
> That malloc library may be doing MADV_DONTNEED calls on memory
> that is no longer needed, expecting those calls to happen on
> PAGE_SIZE boundaries.
> 

Thanks Rik.  I tend to agree with this direction as it is 'breaking'
current code.  David and I discussed this in this thread,
https://lore.kernel.org/linux-mm/356a4b9a-1f56-ae06-b211-bd32fc93ecda@redhat.com/

One thing to note is that there was not any documentation saying
madvise would happen on page boundaries.  The system call takes a
length and rounds up to page size.  However, the man page explicitly
said it operates on a byte range.  Certainly mm people and others
know we only operate on pages.  But, that is not what was documented.

When the change was made to add hugetlb support, the decision was made
to round up the range to hugetlb page boundaries in hugetlb vmas.  This
was to be consistent with how madvise operated on base pages.  At the
same time, madvise documentation was updated say it operates on page
boundaries as well as the behavior for hugetlb mappings.  If moving
forward with this change we will need to update the man page.
  
Andrew Morton Oct. 21, 2022, 8:48 p.m. UTC | #2
On Fri, 21 Oct 2022 15:45:46 -0400 Rik van Riel <riel@surriel.com> wrote:

> A common use case for hugetlbfs is for the application to create
> memory pools backed by huge pages, which then get handed over to
> some malloc library (eg. jemalloc) for further management.
> 
> That malloc library may be doing MADV_DONTNEED calls on memory
> that is no longer needed, expecting those calls to happen on
> PAGE_SIZE boundaries.
> 
> However, currently the MADV_DONTNEED code rounds up any such
> requests to HPAGE_PMD_SIZE boundaries.

Well that's obnoxious.

> This leads to undesired
> outcomes when jemalloc expects a 4kB MADV_DONTNEED, but 2MB of
> memory get zeroed out, instead.
> 
> Use of pre-built shared libraries means that user code does not
> always know the page size of every memory arena in use.
> 
> Avoid unexpected data loss with MADV_DONTNEED by rounding up
> only to PAGE_SIZE (in do_madvise), and rounding down to huge
> page granularity.
> 
> That way programs will only get as much memory zeroed out as
> they requested.

If we merge this, we're inviting people to develop and test code on the 6.2
kernel only to ship it and then find that it misbehaves on 6.1 and
earlier.

So I think we should backport this.

> While we're here, refactor madvise_dontneed_free_valid_vma
> a little so mlocked hugetlb VMAs need MADV_DONTNEED_LOCKED.

And if we do backport it, "while we're here" changes are unwelcome!
  
Rik van Riel Oct. 21, 2022, 11:29 p.m. UTC | #3
On Fri, 2022-10-21 at 13:48 -0700, Mike Kravetz wrote:
> On 10/21/22 15:45, Rik van Riel wrote:
> > A common use case for hugetlbfs is for the application to create
> > memory pools backed by huge pages, which then get handed over to
> > some malloc library (eg. jemalloc) for further management.
> > 
> > That malloc library may be doing MADV_DONTNEED calls on memory
> > that is no longer needed, expecting those calls to happen on
> > PAGE_SIZE boundaries.
> > 
> 
> Thanks Rik.  I tend to agree with this direction as it is 'breaking'
> current code.  David and I discussed this in this thread,
> https://lore.kernel.org/linux-mm/356a4b9a-1f56-ae06-b211-bd32fc93ecda@redhat.com/
> 
> One thing to note is that there was not any documentation saying
> madvise would happen on page boundaries.  The system call takes a
> length and rounds up to page size.  However, the man page explicitly
> said it operates on a byte range.  Certainly mm people and others
> know we only operate on pages.  But, that is not what was documented.
> 
> When the change was made to add hugetlb support, the decision was
> made
> to round up the range to hugetlb page boundaries in hugetlb vmas. 
> This
> was to be consistent with how madvise operated on base pages.  At the
> same time, madvise documentation was updated say it operates on page
> boundaries as well as the behavior for hugetlb mappings.  If moving
> forward with this change we will need to update the man page.

I'll send in a patch for the man page after the patch gets
merged. I'll change the text to clarify that the system
may round up the specified length to PAGE_SIZE granularity,
which is a quantity programs can get through (IIRC) getconf.

Andrew, I split out the bit of the patch for stable.
  
Mike Kravetz Oct. 21, 2022, 11:42 p.m. UTC | #4
On 10/21/22 19:29, Rik van Riel wrote:
> On Fri, 2022-10-21 at 13:48 -0700, Mike Kravetz wrote:
> > On 10/21/22 15:45, Rik van Riel wrote:
> > > A common use case for hugetlbfs is for the application to create
> > > memory pools backed by huge pages, which then get handed over to
> > > some malloc library (eg. jemalloc) for further management.
> > > 
> > > That malloc library may be doing MADV_DONTNEED calls on memory
> > > that is no longer needed, expecting those calls to happen on
> > > PAGE_SIZE boundaries.
> > > 
> > 
> > Thanks Rik.  I tend to agree with this direction as it is 'breaking'
> > current code.  David and I discussed this in this thread,
> > https://lore.kernel.org/linux-mm/356a4b9a-1f56-ae06-b211-bd32fc93ecda@redhat.com/
> > 
> > One thing to note is that there was not any documentation saying
> > madvise would happen on page boundaries.  The system call takes a
> > length and rounds up to page size.  However, the man page explicitly
> > said it operates on a byte range.  Certainly mm people and others
> > know we only operate on pages.  But, that is not what was documented.
> > 
> > When the change was made to add hugetlb support, the decision was
> > made
> > to round up the range to hugetlb page boundaries in hugetlb vmas. 
> > This
> > was to be consistent with how madvise operated on base pages.  At the
> > same time, madvise documentation was updated say it operates on page
> > boundaries as well as the behavior for hugetlb mappings.  If moving
> > forward with this change we will need to update the man page.
> 
> I'll send in a patch for the man page after the patch gets
> merged. I'll change the text to clarify that the system
> may round up the specified length to PAGE_SIZE granularity,
> which is a quantity programs can get through (IIRC) getconf.

Ok, the man page now says madvise only operates on page granularity.
Perhaps we can add info about using sysconf(PAGE_SIZE) if that will
help.  What will really need to be changed is the description of
rounding hugetlb mappings up to the next huge page size boundary.
Just need to describe the hugetlb behavior.  Unfortunately, we may
need to define/describe the arithmetic:
	ALIGN_DOWN(PAGE_ALIGN(length))
i.e. Will align down except in the case length is within PAGE_SIZE
     of hugetlb page size.
  

Patch

diff --git a/mm/madvise.c b/mm/madvise.c
index 2baa93ca2310..a60e8e23c323 100644
--- a/mm/madvise.c
+++ b/mm/madvise.c
@@ -799,21 +799,29 @@  static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
 					    unsigned long *end,
 					    int behavior)
 {
-	if (!is_vm_hugetlb_page(vma)) {
-		unsigned int forbidden = VM_PFNMAP;
+	unsigned int forbidden = VM_PFNMAP;
 
-		if (behavior != MADV_DONTNEED_LOCKED)
-			forbidden |= VM_LOCKED;
+	if (behavior != MADV_DONTNEED_LOCKED)
+		forbidden |= VM_LOCKED;
 
-		return !(vma->vm_flags & forbidden);
-	}
+	if (vma->vm_flags & forbidden)
+		return false;
+
+	if (!is_vm_hugetlb_page(vma))
+		return true;
 
 	if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
 		return false;
 	if (start & ~huge_page_mask(hstate_vma(vma)))
 		return false;
 
-	*end = ALIGN(*end, huge_page_size(hstate_vma(vma)));
+	/*
+	 * Madvise callers expect the length to be rounded up to the page
+	 * size, but they may not know the page size for this VMA is larger
+	 * than PAGE_SIZE! Round down huge pages to avoid unexpected data loss.
+	 */
+	*end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
+
 	return true;
 }
 
@@ -828,6 +836,10 @@  static long madvise_dontneed_free(struct vm_area_struct *vma,
 	if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
 		return -EINVAL;
 
+	/* A small MADV_DONTNEED on a huge page gets rounded down to zero. */
+	if (start == end)
+		return 0;
+
 	if (!userfaultfd_remove(vma, start, end)) {
 		*prev = NULL; /* mmap_lock has been dropped, prev is stale */