[PATCH] x86/fault: Fix wrong signal when vsyscall fails with pkey

[PATCH] x86/fault: Fix wrong signal when vsyscall fails with pkey

[Jiashuo Liang]

The function __bad_area_nosemaphore calls kernelmode_fixup_or_oops with
parameter "signal" being "pkey", which will send a signal numbered "pkey".

When emulating vsyscall, the kernel may fail to access user-given memory
pages that are protected by pkey. In such a case, the kernel should send a
SIGSEGV signal with si_code=SEGV_PKUERR and si_pkey=pkey.

So a new parameter "pkey" is added to kernelmode_fixup_or_oops to fix it.

Signed-off-by: Jiashuo Liang <liangjs@pku.edu.cn>
---
 arch/x86/mm/fault.c | 23 +++++++++++++++--------
 1 file changed, 15 insertions(+), 8 deletions(-)

diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
index b2eefdefc108..883294282e1e 100644
--- a/arch/x86/mm/fault.c
+++ b/arch/x86/mm/fault.c
@@ -710,7 +710,8 @@ page_fault_oops(struct pt_regs *regs, unsigned long error_code,
 
 static noinline void
 kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
-			 unsigned long address, int signal, int si_code)
+			 unsigned long address, int signal, int si_code,
+			 u32 pkey)
 {
 	WARN_ON_ONCE(user_mode(regs));
 
@@ -735,8 +736,12 @@ kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
 
 			set_signal_archinfo(address, error_code);
 
-			/* XXX: hwpoison faults will set the wrong code. */
-			force_sig_fault(signal, si_code, (void __user *)address);
+			if (si_code == SEGV_PKUERR) {
+				force_sig_pkuerr((void __user *)address, pkey);
+			} else {
+				/* XXX: hwpoison faults will set the wrong code. */
+				force_sig_fault(signal, si_code, (void __user *)address);
+			}
 		}
 
 		/*
@@ -798,7 +803,8 @@ __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 	struct task_struct *tsk = current;
 
 	if (!user_mode(regs)) {
-		kernelmode_fixup_or_oops(regs, error_code, address, pkey, si_code);
+		kernelmode_fixup_or_oops(regs, error_code, address,
+					 SIGSEGV, si_code, pkey);
 		return;
 	}
 
@@ -930,7 +936,8 @@ do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
 {
 	/* Kernel mode? Handle exceptions or die: */
 	if (!user_mode(regs)) {
-		kernelmode_fixup_or_oops(regs, error_code, address, SIGBUS, BUS_ADRERR);
+		kernelmode_fixup_or_oops(regs, error_code, address,
+					 SIGBUS, BUS_ADRERR, 0);
 		return;
 	}
 
@@ -1396,7 +1403,7 @@ void do_user_addr_fault(struct pt_regs *regs,
 		 */
 		if (!user_mode(regs))
 			kernelmode_fixup_or_oops(regs, error_code, address,
-						 SIGBUS, BUS_ADRERR);
+						 SIGBUS, BUS_ADRERR, 0);
 		return;
 	}
 
@@ -1416,7 +1423,7 @@ void do_user_addr_fault(struct pt_regs *regs,
 		return;
 
 	if (fatal_signal_pending(current) && !user_mode(regs)) {
-		kernelmode_fixup_or_oops(regs, error_code, address, 0, 0);
+		kernelmode_fixup_or_oops(regs, error_code, address, 0, 0, 0);
 		return;
 	}
 
@@ -1424,7 +1431,7 @@ void do_user_addr_fault(struct pt_regs *regs,
 		/* Kernel mode? Handle exceptions or die: */
 		if (!user_mode(regs)) {
 			kernelmode_fixup_or_oops(regs, error_code, address,
-						 SIGSEGV, SEGV_MAPERR);
+						 SIGSEGV, SEGV_MAPERR, 0);
 			return;
 		}
 
-- 
2.32.0

Re: [PATCH] x86/fault: Fix wrong signal when vsyscall fails with pkey

[Dave Hansen]

On 7/28/21 4:44 AM, Jiashuo Liang wrote:
> The function __bad_area_nosemaphore calls kernelmode_fixup_or_oops with
> parameter "signal" being "pkey", which will send a signal numbered "pkey".

Yikes.

> When emulating vsyscall, the kernel may fail to access user-given memory
> pages that are protected by pkey. In such a case, the kernel should send a
> SIGSEGV signal with si_code=SEGV_PKUERR and si_pkey=pkey.

This could use a bit more context.

First of all this is for user address space faults in the
do_user_addr_fault() path.  Second, the buggy code is under a
!user_mode() check, so this must be a kernel fault in the user address
space.  Third, the only notice this problem when the page fault handler
ends up delivering a signal as a result of the fault.  Most cases will
simply return an error code to the faulting kernel code which will see
-EFAULT come back from copy_to/from_user() and friends.

The *only* condition in which we generate that signal from the fault
handler is when current->thread.sig_on_uaccess_err=1, and the only place
that gets used is in emulate_vsyscall().

This makes me want to add some code that tickles vsyscall emulation in
the pkey selftests, but I think I'll resist the urge for now. :)

Is that all correct?

> So a new parameter "pkey" is added to kernelmode_fixup_or_oops to fix it.

Yeah, I think that's the right fix.  You also need this:

Fixes: 5042d40a264c ("x86/fault: Bypass no_context() for implicit kernel
faults from usermode")

I believe that's where this issue originated.

How did you find this, by the way?

>  arch/x86/mm/fault.c | 23 +++++++++++++++--------
>  1 file changed, 15 insertions(+), 8 deletions(-)
> 
> diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
> index b2eefdefc108..883294282e1e 100644
> --- a/arch/x86/mm/fault.c
> +++ b/arch/x86/mm/fault.c
> @@ -710,7 +710,8 @@ page_fault_oops(struct pt_regs *regs, unsigned long error_code,
>  
>  static noinline void
>  kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
> -			 unsigned long address, int signal, int si_code)
> +			 unsigned long address, int signal, int si_code,
> +			 u32 pkey)
>  {
>  	WARN_ON_ONCE(user_mode(regs));
>  
> @@ -735,8 +736,12 @@ kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
>  
>  			set_signal_archinfo(address, error_code);
>  
> -			/* XXX: hwpoison faults will set the wrong code. */
> -			force_sig_fault(signal, si_code, (void __user *)address);
> +			if (si_code == SEGV_PKUERR) {
> +				force_sig_pkuerr((void __user *)address, pkey);
> +			} else {
> +				/* XXX: hwpoison faults will set the wrong code. */
> +				force_sig_fault(signal, si_code, (void __user *)address);
> +			}
>  		}
>  
>  		/*
> @@ -798,7 +803,8 @@ __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
>  	struct task_struct *tsk = current;
>  
>  	if (!user_mode(regs)) {
> -		kernelmode_fixup_or_oops(regs, error_code, address, pkey, si_code);
> +		kernelmode_fixup_or_oops(regs, error_code, address,
> +					 SIGSEGV, si_code, pkey);
>  		return;
>  	}
>  
> @@ -930,7 +936,8 @@ do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
>  {
>  	/* Kernel mode? Handle exceptions or die: */
>  	if (!user_mode(regs)) {
> -		kernelmode_fixup_or_oops(regs, error_code, address, SIGBUS, BUS_ADRERR);
> +		kernelmode_fixup_or_oops(regs, error_code, address,
> +					 SIGBUS, BUS_ADRERR, 0);
>  		return;
>  	}

Could we please use ARCH_DEFAULT_PKEY instead of 0's in all these call
sites?  I just detest seeing mystery functions with lots of 0's and 1's
as parameters.

Re: [PATCH] x86/fault: Fix wrong signal when vsyscall fails with pkey

[Jiashuo Liang]

On Wed, 2021-07-28 at 10:57 -0700, Dave Hansen wrote:
>> When emulating vsyscall, the kernel may fail to access user-given memory
>> pages that are protected by pkey. In such a case, the kernel should send a
>> SIGSEGV signal with si_code=SEGV_PKUERR and si_pkey=pkey.
> 
> This could use a bit more context.
> 
> First of all this is for user address space faults in the
> do_user_addr_fault() path.  Second, the buggy code is under a
> !user_mode() check, so this must be a kernel fault in the user address
> space.  Third, the only notice this problem when the page fault handler
> ends up delivering a signal as a result of the fault.  Most cases will
> simply return an error code to the faulting kernel code which will see
> -EFAULT come back from copy_to/from_user() and friends.
> 
> The *only* condition in which we generate that signal from the fault
> handler is when current->thread.sig_on_uaccess_err=1, and the only place
> that gets used is in emulate_vsyscall().
> 
> This makes me want to add some code that tickles vsyscall emulation in
> the pkey selftests, but I think I'll resist the urge for now. :)
> 
> Is that all correct?

Right.

>> So a new parameter "pkey" is added to kernelmode_fixup_or_oops to fix it.
> 
> Yeah, I think that's the right fix.  You also need this:
> 
> Fixes: 5042d40a264c ("x86/fault: Bypass no_context() for implicit kernel
> faults from usermode")
> 
> I believe that's where this issue originated.

Yeah, we need to add it.

> How did you find this, by the way?

I was learning about memory protection key. So I read the related code in
kernel and spotted this.

>>  arch/x86/mm/fault.c | 23 +++++++++++++++--------
>>  1 file changed, 15 insertions(+), 8 deletions(-)
>> 
>> diff --git a/arch/x86/mm/fault.c b/arch/x86/mm/fault.c
>> index b2eefdefc108..883294282e1e 100644
>> --- a/arch/x86/mm/fault.c
>> +++ b/arch/x86/mm/fault.c
>> @@ -710,7 +710,8 @@ page_fault_oops(struct pt_regs *regs, unsigned long error_code,
>>  
>>  static noinline void
>>  kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
>> -			 unsigned long address, int signal, int si_code)
>> +			 unsigned long address, int signal, int si_code,
>> +			 u32 pkey)
>>  {
>>  	WARN_ON_ONCE(user_mode(regs));
>>  
>> @@ -735,8 +736,12 @@ kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
>>  
>>  			set_signal_archinfo(address, error_code);
>>  
>> -			/* XXX: hwpoison faults will set the wrong code. */
>> -			force_sig_fault(signal, si_code, (void __user *)address);
>> +			if (si_code == SEGV_PKUERR) {
>> +				force_sig_pkuerr((void __user *)address, pkey);
>> +			} else {
>> +				/* XXX: hwpoison faults will set the wrong code. */
>> +				force_sig_fault(signal, si_code, (void __user *)address);
>> +			}
>>  		}
>>  
>>  		/*
>> @@ -798,7 +803,8 @@ __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
>>  	struct task_struct *tsk = current;
>>  
>>  	if (!user_mode(regs)) {
>> -		kernelmode_fixup_or_oops(regs, error_code, address, pkey, si_code);
>> +		kernelmode_fixup_or_oops(regs, error_code, address,
>> +					 SIGSEGV, si_code, pkey);
>>  		return;
>>  	}
>>  
>> @@ -930,7 +936,8 @@ do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
>>  {
>>  	/* Kernel mode? Handle exceptions or die: */
>>  	if (!user_mode(regs)) {
>> -		kernelmode_fixup_or_oops(regs, error_code, address, SIGBUS, BUS_ADRERR);
>> +		kernelmode_fixup_or_oops(regs, error_code, address,
>> +					 SIGBUS, BUS_ADRERR, 0);
>>  		return;
>>  	}
> 
> Could we please use ARCH_DEFAULT_PKEY instead of 0's in all these call
> sites?  I just detest seeing mystery functions with lots of 0's and 1's
> as parameters.

I agree that using ARCH_DEFAULT_PKEY is better. I think I am supposed to
send a patch v2 for the update?

By the way, the magic pkey number 0 also appears when bad_area_nosemaphore
calls __bad_area_nosemaphore and bad_area calls __bad_area. Do they need to
be changed to ARCH_DEFAULT_PKEY as well?

Thanks!

Jiashuo Liang

Re: [PATCH] x86/fault: Fix wrong signal when vsyscall fails with pkey

[Dave Hansen]

On 7/28/21 11:24 PM, Jiashuo Liang wrote:
> I agree that using ARCH_DEFAULT_PKEY is better. I think I am supposed to
> send a patch v2 for the update?

Yes, please.

> By the way, the magic pkey number 0 also appears when bad_area_nosemaphore
> calls __bad_area_nosemaphore and bad_area calls __bad_area. Do they need to
> be changed to ARCH_DEFAULT_PKEY as well?

Yes, it would be best to change those as well.  Ideally in a separate patch.

Re: [PATCH] x86/fault: Fix wrong signal when vsyscall fails with pkey

[Jiashuo Liang]

On Thu, 2021-07-29 at 07:02 -0700, Dave Hansen wrote:
> On 7/28/21 11:24 PM, Jiashuo Liang wrote:
>> I agree that using ARCH_DEFAULT_PKEY is better. I think I am supposed to
>> send a patch v2 for the update?
>
> Yes, please.

The patch v2 is here: https://lore.kernel.org/lkml/20210730030152.249106-1-liangjs@pku.edu.cn/
Would you please have a look?