[PATCH v2] x86/mm/fault: Allow stack access below %rsp

[PATCH v2] x86/mm/fault: Allow stack access below %rsp

[Waiman Long]

The current x86 page fault handler allows stack access below the stack
pointer if it is no more than 64k+256 bytes. Any access beyond the 64k+
limit will cause a segmentation fault.

The gcc -fstack-check option generates code to probe the stack for
large stack allocation to see if the stack is accessible. The newer gcc
does that while updating the %rsp simultaneously. Older gcc's like gcc4
doesn't do that. As a result, an application compiled with an old gcc
and the -fstack-check option may fail to start at all.

% cat test.c
int main() {
	char tmp[1024*128];
	printf("### ok\n");
	return 0;
}
% gcc -fstack-check -g -o test test.c
% ./test
Segmentation fault

The old binary was working in older kernels where expand_stack() was
somehow called before the check. But it is not working in newer kernels.
Besides, the 64k+ limit check is kind of crude and will not catch a
lot of mistakes that userspace applications may be misbehaving anyway.
I think the kernel isn't the right place for this kind of tests. We
should leave it to userspace instrumentation tools to perform them.

The 64k+ limit check is now removed to just let expand_stack() decide
if a segmentation fault should happen, when the RLIMIT_STACK limit is
exceeded, for example.

Signed-off-by: Waiman Long <longman@redhat.com>
---
 arch/x86/mm/fault.c | 12 ------------
 1 file changed, 12 deletions(-)

diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
index 71d4b9d..29525cf 100644
--- a/arch/x86/mm/fault.c
+++ b/arch/x86/mm/fault.c
@@ -1380,18 +1380,6 @@ void do_user_addr_fault(struct pt_regs *regs,
 		bad_area(regs, sw_error_code, address);
 		return;
 	}
-	if (sw_error_code & X86_PF_USER) {
-		/*
-		 * Accessing the stack below %sp is always a bug.
-		 * The large cushion allows instructions like enter
-		 * and pusha to work. ("enter $65535, $31" pushes
-		 * 32 pointers and then decrements %sp by 65535.)
-		 */
-		if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
-			bad_area(regs, sw_error_code, address);
-			return;
-		}
-	}
 	if (unlikely(expand_stack(vma, address))) {
 		bad_area(regs, sw_error_code, address);
 		return;
-- 
1.8.3.1

Re: [PATCH v2] x86/mm/fault: Allow stack access below %rsp

[Waiman Long]

On 11/06/2018 03:12 PM, Waiman Long wrote:
> The current x86 page fault handler allows stack access below the stack
> pointer if it is no more than 64k+256 bytes. Any access beyond the 64k+
> limit will cause a segmentation fault.
>
> The gcc -fstack-check option generates code to probe the stack for
> large stack allocation to see if the stack is accessible. The newer gcc
> does that while updating the %rsp simultaneously. Older gcc's like gcc4
> doesn't do that. As a result, an application compiled with an old gcc
> and the -fstack-check option may fail to start at all.
>
> % cat test.c
> int main() {
> 	char tmp[1024*128];
> 	printf("### ok\n");
> 	return 0;
> }
> % gcc -fstack-check -g -o test test.c
> % ./test
> Segmentation fault
>
> The old binary was working in older kernels where expand_stack() was
> somehow called before the check. But it is not working in newer kernels.
> Besides, the 64k+ limit check is kind of crude and will not catch a
> lot of mistakes that userspace applications may be misbehaving anyway.
> I think the kernel isn't the right place for this kind of tests. We
> should leave it to userspace instrumentation tools to perform them.
>
> The 64k+ limit check is now removed to just let expand_stack() decide
> if a segmentation fault should happen, when the RLIMIT_STACK limit is
> exceeded, for example.
>
> Signed-off-by: Waiman Long <longman@redhat.com>
> ---
>  arch/x86/mm/fault.c | 12 ------------
>  1 file changed, 12 deletions(-)
>
> diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
> index 71d4b9d..29525cf 100644
> --- a/arch/x86/mm/fault.c
> +++ b/arch/x86/mm/fault.c
> @@ -1380,18 +1380,6 @@ void do_user_addr_fault(struct pt_regs *regs,
>  		bad_area(regs, sw_error_code, address);
>  		return;
>  	}
> -	if (sw_error_code & X86_PF_USER) {
> -		/*
> -		 * Accessing the stack below %sp is always a bug.
> -		 * The large cushion allows instructions like enter
> -		 * and pusha to work. ("enter $65535, $31" pushes
> -		 * 32 pointers and then decrements %sp by 65535.)
> -		 */
> -		if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
> -			bad_area(regs, sw_error_code, address);
> -			return;
> -		}
> -	}
>  	if (unlikely(expand_stack(vma, address))) {
>  		bad_area(regs, sw_error_code, address);
>  		return;

This v2 patch has no code change. I just updated the commit log to
capture some of the conversion that I had with the reviewers.

Cheers,
Longman

[tip:x86/mm] x86/mm/fault: Allow stack access below %rsp

[tip-bot for Waiman Long]

Commit-ID:  1d8ca3be86ebc6a38dad8236f45c7a9c61681e78
Gitweb:     https://git.kernel.org/tip/1d8ca3be86ebc6a38dad8236f45c7a9c61681e78
Author:     Waiman Long <longman@redhat.com>
AuthorDate: Tue, 6 Nov 2018 15:12:29 -0500
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Mon, 12 Nov 2018 11:06:19 +0100

x86/mm/fault: Allow stack access below %rsp

The current x86 page fault handler allows stack access below the stack
pointer if it is no more than 64k+256 bytes. Any access beyond the 64k+
limit will cause a segmentation fault.

The gcc -fstack-check option generates code to probe the stack for
large stack allocation to see if the stack is accessible. The newer gcc
does that while updating the %rsp simultaneously. Older gcc's like gcc4
doesn't do that. As a result, an application compiled with an old gcc
and the -fstack-check option may fail to start at all:

  $ cat test.c
  int main() {
	char tmp[1024*128];
	printf("### ok\n");
	return 0;
  }

  $ gcc -fstack-check -g -o test test.c

  $ ./test
  Segmentation fault

The old binary was working in older kernels where expand_stack() was
somehow called before the check. But it is not working in newer kernels.
Besides, the 64k+ limit check is kind of crude and will not catch a
lot of mistakes that userspace applications may be misbehaving anyway.
I think the kernel isn't the right place for this kind of tests. We
should leave it to userspace instrumentation tools to perform them.

The 64k+ limit check is now removed to just let expand_stack() decide
if a segmentation fault should happen, when the RLIMIT_STACK limit is
exceeded, for example.

Signed-off-by: Waiman Long <longman@redhat.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1541535149-31963-1-git-send-email-longman@redhat.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 arch/x86/mm/fault.c | 12 ------------
 1 file changed, 12 deletions(-)

diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
index 71d4b9d4d43f..29525cf21100 100644
--- a/arch/x86/mm/fault.c
+++ b/arch/x86/mm/fault.c
@@ -1380,18 +1380,6 @@ retry:
 		bad_area(regs, sw_error_code, address);
 		return;
 	}
-	if (sw_error_code & X86_PF_USER) {
-		/*
-		 * Accessing the stack below %sp is always a bug.
-		 * The large cushion allows instructions like enter
-		 * and pusha to work. ("enter $65535, $31" pushes
-		 * 32 pointers and then decrements %sp by 65535.)
-		 */
-		if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
-			bad_area(regs, sw_error_code, address);
-			return;
-		}
-	}
 	if (unlikely(expand_stack(vma, address))) {
 		bad_area(regs, sw_error_code, address);
 		return;