[RFC PATCH] Take mmap_min_addr into account while choosing unmapped address for x86-64.

[RFC PATCH] Take mmap_min_addr into account while choosing unmapped address for x86-64.

[Ilya Smith]

It prevent situation when vm_unmapped_area chose address between
PAGE_SIZE and mmap_min_addr range. In this case mmap will fail with
EPERM without a good reason.

As test-case of such situation we may hard-code address between
PAGE_SIZE and 65536 inside unmapped_area_topdown function.

Signed-off-by: Ilya Smith <blackzert@gmail.com>
---
 arch/x86/kernel/sys_x86_64.c | 5 +++--
 arch/x86/mm/mmap.c           | 4 ++++
 2 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kernel/sys_x86_64.c b/arch/x86/kernel/sys_x86_64.c
index 676774b9bb8d..1752fe5cb735 100644
--- a/arch/x86/kernel/sys_x86_64.c
+++ b/arch/x86/kernel/sys_x86_64.c
@@ -17,6 +17,7 @@
 #include <linux/random.h>
 #include <linux/uaccess.h>
 #include <linux/elf.h>
+#include <linux/security.h>
 
 #include <asm/elf.h>
 #include <asm/compat.h>
@@ -185,7 +186,7 @@ arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
 		return addr;
 
 	/* requested length too big for entire address space */
-	if (len > TASK_SIZE)
+	if (len > TASK_SIZE - mmap_min_addr)
 		return -ENOMEM;
 
 	/* No address checking. See comment at mmap_address_hint_valid() */
@@ -210,7 +211,7 @@ arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
 
 	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 	info.length = len;
-	info.low_limit = PAGE_SIZE;
+	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
 	info.high_limit = get_mmap_base(0);
 
 	/*
diff --git a/arch/x86/mm/mmap.c b/arch/x86/mm/mmap.c
index 155ecbac9e28..b6d0c317639e 100644
--- a/arch/x86/mm/mmap.c
+++ b/arch/x86/mm/mmap.c
@@ -31,6 +31,7 @@
 #include <linux/sched/signal.h>
 #include <linux/sched/mm.h>
 #include <linux/compat.h>
+#include <linux/security.h>
 #include <asm/elf.h>
 
 #include "physaddr.h"
@@ -220,6 +221,9 @@ bool mmap_address_hint_valid(unsigned long addr, unsigned long len)
 	if (TASK_SIZE - len < addr)
 		return false;
 
+	if (addr < mmap_min_addr)
+		return false;
+
 	return (addr > DEFAULT_MAP_WINDOW) == (addr + len > DEFAULT_MAP_WINDOW);
 }
 
-- 
2.14.1

Re: [RFC PATCH] Take mmap_min_addr into account while choosing unmapped address for x86-64.

[Michal Hocko]

On Tue 27-02-18 00:12:57, Ilya Smith wrote:
[...]
> diff --git a/arch/x86/kernel/sys_x86_64.c b/arch/x86/kernel/sys_x86_64.c
> index 676774b9bb8d..1752fe5cb735 100644
> --- a/arch/x86/kernel/sys_x86_64.c
> +++ b/arch/x86/kernel/sys_x86_64.c
> @@ -17,6 +17,7 @@
>  #include <linux/random.h>
>  #include <linux/uaccess.h>
>  #include <linux/elf.h>
> +#include <linux/security.h>
>  
>  #include <asm/elf.h>
>  #include <asm/compat.h>
> @@ -185,7 +186,7 @@ arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
>  		return addr;
>  
>  	/* requested length too big for entire address space */
> -	if (len > TASK_SIZE)
> +	if (len > TASK_SIZE - mmap_min_addr)
>  		return -ENOMEM;
>  
>  	/* No address checking. See comment at mmap_address_hint_valid() */
> @@ -210,7 +211,7 @@ arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
>  
>  	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
>  	info.length = len;
> -	info.low_limit = PAGE_SIZE;
> +	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
>  	info.high_limit = get_mmap_base(0);
>  
>  	/*

mmap_min_addr handling is a bit mess... As you say, we would return
EPERM rather than ENOMEM which can be confusing but depleting the
address space like that is quite unlikely on 64b unless I am missing.
It is good to be in sync here with the generic implementation though,
IMO.

> diff --git a/arch/x86/mm/mmap.c b/arch/x86/mm/mmap.c
> index 155ecbac9e28..b6d0c317639e 100644
> --- a/arch/x86/mm/mmap.c
> +++ b/arch/x86/mm/mmap.c
> @@ -31,6 +31,7 @@
>  #include <linux/sched/signal.h>
>  #include <linux/sched/mm.h>
>  #include <linux/compat.h>
> +#include <linux/security.h>
>  #include <asm/elf.h>
>  
>  #include "physaddr.h"
> @@ -220,6 +221,9 @@ bool mmap_address_hint_valid(unsigned long addr, unsigned long len)
>  	if (TASK_SIZE - len < addr)
>  		return false;
>  
> +	if (addr < mmap_min_addr)
> +		return false;
> +
>  	return (addr > DEFAULT_MAP_WINDOW) == (addr + len > DEFAULT_MAP_WINDOW);

But is this one necessary? We do sanitze hint address before going to
get_unmapped_address AFAIR.

-- 
Michal Hocko
SUSE Labs

Re: [RFC PATCH] Take mmap_min_addr into account while choosing unmapped address for x86-64.

[Ilya Smith]

> 
> mmap_min_addr handling is a bit mess... As you say, we would return
> EPERM rather than ENOMEM which can be confusing but depleting the
> address space like that is quite unlikely on 64b unless I am missing.
> It is good to be in sync here with the generic implementation though,
> IMO.
> 

If we take a look on mm/mmap.c:
#ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
unsigned long
arch_get_unmapped_area_topdown(
…
	if (len > TASK_SIZE - mmap_min_addr)
		return -ENOMEM;
…
	info.low_limit = max(PAGE_SIZE, mmap_min_addr);

And this one looks like a generic implementation.
But for many other architectures like arch/parisc/kernel/sys_parisc.c 
or arch/x86/kernel/sys_x86_64.c

	info.low_limit = PAGE_SIZE;

What is looks like an issue for me.

Here is C code could be used as test-case:

#include <errno.h>
#include <stdio.h>
#include <sys/mman.h>

int main() {
  char buffer[1024];

  unsigned long len = 1ULL << 46;
  while(len) {
    void *ptr = mmap(4096, len, 0, MAP_PRIVATE | MAP_ANONYMOUS , -1, 0);
    if (ptr == MAP_FAILED) {
      if (errno == EPERM)
        break;
      if (errno == ENOMEM) {
        len >>= 1;
        continue;
      }
      return -1;
    }
  }
  if (errno == EPERM) {
    printf("Test failed, you have wrong ret code EPERM\n");
    sprintf(buffer, "cat /proc/%d/maps", getpid());
    system(buffer);
    return -1;
  }
  return 0;
}


>> 
>> +	if (addr < mmap_min_addr)
>> +		return false;
>> +
>> 	return (addr > DEFAULT_MAP_WINDOW) == (addr + len > DEFAULT_MAP_WINDOW);
> 
> But is this one necessary? We do sanitze hint address before going to
> get_unmapped_address AFAIR.
> 
I’m agree, looks like I was trying to fix something that already fine.

Thanks,
Ilya

Re: [RFC PATCH] Take mmap_min_addr into account while choosing unmapped address for x86-64.

[Michal Hocko]

On Tue 27-02-18 16:27:29, Ilya Smith wrote:
> > 
> > mmap_min_addr handling is a bit mess... As you say, we would return
> > EPERM rather than ENOMEM which can be confusing but depleting the
> > address space like that is quite unlikely on 64b unless I am missing.
> > It is good to be in sync here with the generic implementation though,
> > IMO.
> > 
> 
> If we take a look on mm/mmap.c:
> #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
> unsigned long
> arch_get_unmapped_area_topdown(
> …
> 	if (len > TASK_SIZE - mmap_min_addr)
> 		return -ENOMEM;
> …
> 	info.low_limit = max(PAGE_SIZE, mmap_min_addr);
> 
> And this one looks like a generic implementation.
> But for many other architectures like arch/parisc/kernel/sys_parisc.c 
> or arch/x86/kernel/sys_x86_64.c
> 
> 	info.low_limit = PAGE_SIZE;

Yeah, this is what I meant when saying that mmap_min_addr is a bit of a
mess. I am wondering whether the low_limit should be checked inside
vm_unmapped_area. We would still need some mmap_min_addr handling at
arch_get_unmapped_area_topdown layer which is still suboptimal but I do
not see an easy way around without reworking how the arch specific parts
are implemented currently.
-- 
Michal Hocko
SUSE Labs