remap_file_pages regression

remap_file_pages regression

[Grazvydas Ignotas]

Hi,

since remap_file_pages() rework the following simple program fails.
I haven't actually bisected this, only know it worked on 3.19 at least
(I bought a new system now and need 4.2+ for hardware support). If you
are curious, the program is an emulator and is using remap_file_pages()
to implement memory mirroring efficiently (and to remap things during
run time).

Grazvydas

====%<===

#define _GNU_SOURCE
#include <assert.h>
#include <stdlib.h>
#include <stdio.h>
#include <sys/mman.h>

#define SIZE	(4096 * 3)

int main(int argc, char **argv)
{
	unsigned long *p;
	long i;

	p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
			MAP_SHARED | MAP_ANONYMOUS, -1, 0);
	if (p == MAP_FAILED) {
		perror("mmap");
		return -1;
	}

	for (i = 0; i < SIZE / 4096; i++)
		p[i * 4096 / sizeof(*p)] = i;

	if (remap_file_pages(p, 4096, 0, 1, 0)) {
		perror("remap_file_pages");
		return -1;
	}

	if (remap_file_pages(p, 4096 * 2, 0, 1, 0)) {
		perror("remap_file_pages");
		return -1;
	}

	assert(p[0] == 1);

	munmap(p, SIZE);

	return 0;
}

Re: remap_file_pages regression

[Kirill A. Shutemov]

On Mon, Feb 15, 2016 at 03:32:55AM +0200, Grazvydas Ignotas wrote:
> Hi,
> 
> since remap_file_pages() rework the following simple program fails.
> I haven't actually bisected this, only know it worked on 3.19 at least
> (I bought a new system now and need 4.2+ for hardware support).

The patch below should fix the issue. Please test.

> If you are curious, the program is an emulator and is using
> remap_file_pages() to implement memory mirroring efficiently (and to
> remap things during run time).

Could you elaborate on this?

Why creating file on tmpfs/shmem (using memfd_create() for example) plus
plain mmap()s wouldn't work for you?

>From 6b8690c3a983a72212db8cd35eb4cab106a195d4 Mon Sep 17 00:00:00 2001
From: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Date: Mon, 15 Feb 2016 13:06:16 +0300
Subject: [PATCH] mm: fix regression in remap_file_pages() emulation

Grazvydas Ignotas has reported a regression in remap_file_pages()
emulation.

Testcase:
	#define _GNU_SOURCE
	#include <assert.h>
	#include <stdlib.h>
	#include <stdio.h>
	#include <sys/mman.h>

	#define SIZE    (4096 * 3)

	int main(int argc, char **argv)
	{
		unsigned long *p;
		long i;

		p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
				MAP_SHARED | MAP_ANONYMOUS, -1, 0);
		if (p == MAP_FAILED) {
			perror("mmap");
			return -1;
		}

		for (i = 0; i < SIZE / 4096; i++)
			p[i * 4096 / sizeof(*p)] = i;

		if (remap_file_pages(p, 4096, 0, 1, 0)) {
			perror("remap_file_pages");
			return -1;
		}

		if (remap_file_pages(p, 4096 * 2, 0, 1, 0)) {
			perror("remap_file_pages");
			return -1;
		}

		assert(p[0] == 1);

		munmap(p, SIZE);

		return 0;
	}

The second remap_file_pages() fails with -EINVAL.

The reason is that remap_file_pages() emulation assumes that the target
vma covers whole area we want to over map. That assumption is broken by
first remap_file_pages() call: it split the area into two vma.

The solution is to check next adjacent vmas, if they map the same file
with the same flags.

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reported-by: Grazvydas Ignotas <notasas@gmail.com>
---
 mm/mmap.c | 34 +++++++++++++++++++++++++++++-----
 1 file changed, 29 insertions(+), 5 deletions(-)

diff --git a/mm/mmap.c b/mm/mmap.c
index 2f2415a7a688..76d1ec29149b 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -2664,12 +2664,29 @@ SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
 	if (!vma || !(vma->vm_flags & VM_SHARED))
 		goto out;
 
-	if (start < vma->vm_start || start + size > vma->vm_end)
+	if (start < vma->vm_start)
 		goto out;
 
-	if (pgoff == linear_page_index(vma, start)) {
-		ret = 0;
-		goto out;
+	if (start + size > vma->vm_end) {
+		struct vm_area_struct *next;
+
+		for (next = vma->vm_next; next; next = next->vm_next) {
+			/* hole between vmas ? */
+			if (next->vm_start != next->vm_prev->vm_end)
+				goto out;
+
+			if (next->vm_file != vma->vm_file)
+				goto out;
+
+			if (next->vm_flags != vma->vm_flags)
+				goto out;
+
+			if (start + size <= next->vm_end)
+				break;
+		}
+
+		if (!next)
+			goto out;
 	}
 
 	prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
@@ -2679,9 +2696,16 @@ SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
 	flags &= MAP_NONBLOCK;
 	flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
 	if (vma->vm_flags & VM_LOCKED) {
+		struct vm_area_struct *tmp;
 		flags |= MAP_LOCKED;
+
 		/* drop PG_Mlocked flag for over-mapped range */
-		munlock_vma_pages_range(vma, start, start + size);
+		for (tmp = vma; tmp->vm_start >= start + size;
+				tmp = tmp->vm_next) {
+			munlock_vma_pages_range(tmp,
+					max(tmp->vm_start, start),
+					min(tmp->vm_end, start + size));
+		}
 	}
 
 	file = get_file(vma->vm_file);
-- 
 Kirill A. Shutemov

Re: remap_file_pages regression

[Grazvydas Ignotas]

On Mon, Feb 15, 2016 at 12:26 PM, Kirill A. Shutemov
<kirill.shutemov@linux.intel.com> wrote:
> On Mon, Feb 15, 2016 at 03:32:55AM +0200, Grazvydas Ignotas wrote:
>> Hi,
>>
>> since remap_file_pages() rework the following simple program fails.
>> I haven't actually bisected this, only know it worked on 3.19 at least
>> (I bought a new system now and need 4.2+ for hardware support).
>
> The patch below should fix the issue. Please test.

It works, thanks.
Tested-by: Grazvydas Ignotas <notasas@gmail.com>

>
>> If you are curious, the program is an emulator and is using
>> remap_file_pages() to implement memory mirroring efficiently (and to
>> remap things during run time).
>
> Could you elaborate on this?
>
> Why creating file on tmpfs/shmem (using memfd_create() for example) plus
> plain mmap()s wouldn't work for you?

It works, but remap_file_pages() is just more convenient, you don't
need many mmap()/munmap() calls (less syscalls), and as the emulator
needs to reconfigure the mappings during runtime (it's not a one time
init thing), remap_file_pages() makes more sense. The reduced number
of VMAs of the past was also a benefit I guess.

Actually I'm not the author of the emulator in question, so I've asked
the author to comment:
--- quote ---
One of the things we do with remap_file_pages is a mapping for the
emulated system's VRAM.  The system allows many different
configurations of several differently sized VRAM banks to a 16MB area
of address space. The banks are multiples of 16KB in size and their
allocations are also 16KB aligned. With remap_file_pages we can map
that entire space to an (in-memory) file and then arbitrarily map
chunks of it as desired. We have an mmap-based fallback, but it
requires us to manage a separate mmap for each 16KB region."
--- end of quote ---
Apparently the mentioned fallback was a compile time option, so I was
hit with the issue after upgrading my machine.

> <snip>
>
> Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

Cc: stable?
Fixes: c8d78c182 (I guess?)

> Reported-by: Grazvydas Ignotas <notasas@gmail.com>
> ---
>  mm/mmap.c | 34 +++++++++++++++++++++++++++++-----
>  1 file changed, 29 insertions(+), 5 deletions(-)
>
> diff --git a/mm/mmap.c b/mm/mmap.c
> index 2f2415a7a688..76d1ec29149b 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -2664,12 +2664,29 @@ SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
>         if (!vma || !(vma->vm_flags & VM_SHARED))
>                 goto out;
>
> -       if (start < vma->vm_start || start + size > vma->vm_end)
> +       if (start < vma->vm_start)
>                 goto out;
>
> -       if (pgoff == linear_page_index(vma, start)) {
> -               ret = 0;
> -               goto out;
> +       if (start + size > vma->vm_end) {
> +               struct vm_area_struct *next;
> +
> +               for (next = vma->vm_next; next; next = next->vm_next) {
> +                       /* hole between vmas ? */
> +                       if (next->vm_start != next->vm_prev->vm_end)
> +                               goto out;
> +
> +                       if (next->vm_file != vma->vm_file)
> +                               goto out;
> +
> +                       if (next->vm_flags != vma->vm_flags)
> +                               goto out;
> +
> +                       if (start + size <= next->vm_end)
> +                               break;
> +               }
> +
> +               if (!next)
> +                       goto out;
>         }
>
>         prot |= vma->vm_flags & VM_READ ? PROT_READ : 0;
> @@ -2679,9 +2696,16 @@ SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
>         flags &= MAP_NONBLOCK;
>         flags |= MAP_SHARED | MAP_FIXED | MAP_POPULATE;
>         if (vma->vm_flags & VM_LOCKED) {
> +               struct vm_area_struct *tmp;
>                 flags |= MAP_LOCKED;
> +
>                 /* drop PG_Mlocked flag for over-mapped range */
> -               munlock_vma_pages_range(vma, start, start + size);
> +               for (tmp = vma; tmp->vm_start >= start + size;
> +                               tmp = tmp->vm_next) {
> +                       munlock_vma_pages_range(tmp,
> +                                       max(tmp->vm_start, start),
> +                                       min(tmp->vm_end, start + size));
> +               }
>         }
>
>         file = get_file(vma->vm_file);
> --
>  Kirill A. Shutemov

Gražvydas

Re: remap_file_pages regression

[Kirill A. Shutemov]

On Tue, Feb 16, 2016 at 12:08:34AM +0200, Grazvydas Ignotas wrote:
> On Mon, Feb 15, 2016 at 12:26 PM, Kirill A. Shutemov
> <kirill.shutemov@linux.intel.com> wrote:
> > On Mon, Feb 15, 2016 at 03:32:55AM +0200, Grazvydas Ignotas wrote:
> >> Hi,
> >>
> >> since remap_file_pages() rework the following simple program fails.
> >> I haven't actually bisected this, only know it worked on 3.19 at least
> >> (I bought a new system now and need 4.2+ for hardware support).
> >
> > The patch below should fix the issue. Please test.
> 
> It works, thanks.
> Tested-by: Grazvydas Ignotas <notasas@gmail.com>
> 
> >
> >> If you are curious, the program is an emulator and is using
> >> remap_file_pages() to implement memory mirroring efficiently (and to
> >> remap things during run time).
> >
> > Could you elaborate on this?
> >
> > Why creating file on tmpfs/shmem (using memfd_create() for example) plus
> > plain mmap()s wouldn't work for you?
> 
> It works, but remap_file_pages() is just more convenient, you don't
> need many mmap()/munmap() calls (less syscalls)

mmap(MAP_FIXED) would do munmap() for you.

> > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> 
> Cc: stable?
> Fixes: c8d78c182 (I guess?)

Right.

Cc: stable@vger.kernel.org # v4.0+
Fixes: c8d78c1823f4 ("mm: replace remap_file_pages() syscall with emulation")

-- 
 Kirill A. Shutemov