[bpf,v2,2/3] x86/mm: Disallow vsyscall page read for copy_from_kernel_nofault()

Commit Message

Hou Tao Jan. 26, 2024, 11:54 a.m. UTC

From: Hou Tao <[e-mail]>

When trying to use copy_from_kernel_nofault() to read vsyscall page
through a bpf program, the following oops was reported:

  BUG: unable to handle page fault for address: ffffffffff600000
  #PF: supervisor read access in kernel mode
  #PF: error_code(0x0000) - not-present page
  PGD 3231067 P4D 3231067 PUD 3233067 PMD 3235067 PTE 0
  Oops: 0000 [#1] PREEMPT SMP PTI
  CPU: 1 PID: 20390 Comm: test_progs ...... 6.7.0+ #58
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996) ......
  RIP: 0010:copy_from_kernel_nofault+0x6f/0x110
  ......
  Call Trace:
   <TASK>
   ? copy_from_kernel_nofault+0x6f/0x110
   bpf_probe_read_kernel+0x1d/0x50
   bpf_prog_2061065e56845f08_do_probe_read+0x51/0x8d
   trace_call_bpf+0xc5/0x1c0
   perf_call_bpf_enter.isra.0+0x69/0xb0
   perf_syscall_enter+0x13e/0x200
   syscall_trace_enter+0x188/0x1c0
   do_syscall_64+0xb5/0xe0
   entry_SYSCALL_64_after_hwframe+0x6e/0x76
   </TASK>
  ......
  ---[ end trace 0000000000000000 ]---

It seems the occurrence of oops depends on SMAP feature of CPU. It
happens as follow: a bpf program uses bpf_probe_read_kernel() to read
from vsyscall page, bpf_probe_read_kernel() invokes
copy_from_kernel_nofault() in turn and then invokes __get_user_asm().
Because the vsyscall page address is not readable for kernel space,
a page fault exception is triggered accordingly, handle_page_fault()
considers the vsyscall page address as a userspace address instead of a
kernel space address, so the fix-up set-up by bpf isn't applied. Because
the CPU has SMAP feature and the access happens in kernel mode, so
page_fault_oops() is invoked and an oops happens. If these is no SMAP
feature, the fix-up set-up by bpf will be applied and
copy_from_kernel_nofault() will return -EFAULT instead.

Considering handle_page_fault() has already considered the vsyscall page
address as a userspace address, fix the problem by disallowing vsyscall
page read for copy_from_kernel_nofault().

Originally-by: Thomas Gleixner <[e-mail]>
Reported-by: [e-mail]
Closes: https://lore.kernel.org/bpf/CAG48ez06TZft=[e-mail]
Reported-by: xingwei lee <[e-mail]>
Closes: https://lore.kernel.org/bpf/CABOYnLynjBoFZOf3Z4BhaZkc5hx_kHfsjiW+UWLoB=[e-mail]
Signed-off-by: Hou Tao <[e-mail]>
---
 arch/x86/mm/maccess.c | 9 +++++++++
 1 file changed, 9 insertions(+)

Comments

Sohil Mehta Jan. 29, 2024, 11:50 p.m. UTC | #1

Hi Hou Tao,

I agree to your approach in this patch. Please see some comments below.

On 1/26/2024 3:54 AM, Hou Tao wrote:
> From: Hou Tao <[e-mail]>
> 
> When trying to use copy_from_kernel_nofault() to read vsyscall page
> through a bpf program, the following oops was reported:
> 
>   BUG: unable to handle page fault for address: ffffffffff600000
>   #PF: supervisor read access in kernel mode
>   #PF: error_code(0x0000) - not-present page
>   PGD 3231067 P4D 3231067 PUD 3233067 PMD 3235067 PTE 0
>   Oops: 0000 [#1] PREEMPT SMP PTI
>   CPU: 1 PID: 20390 Comm: test_progs ...... 6.7.0+ #58
>   Hardware name: QEMU Standard PC (i440FX + PIIX, 1996) ......
>   RIP: 0010:copy_from_kernel_nofault+0x6f/0x110
>   ......
>   Call Trace:
>    <TASK>
>    ? copy_from_kernel_nofault+0x6f/0x110
>    bpf_probe_read_kernel+0x1d/0x50
>    bpf_prog_2061065e56845f08_do_probe_read+0x51/0x8d
>    trace_call_bpf+0xc5/0x1c0
>    perf_call_bpf_enter.isra.0+0x69/0xb0
>    perf_syscall_enter+0x13e/0x200
>    syscall_trace_enter+0x188/0x1c0
>    do_syscall_64+0xb5/0xe0
>    entry_SYSCALL_64_after_hwframe+0x6e/0x76
>    </TASK>
>   ......
>   ---[ end trace 0000000000000000 ]---
> 


> It seems the occurrence of oops depends on SMAP feature of CPU. It
> happens as follow: a bpf program uses bpf_probe_read_kernel() to read
> from vsyscall page, bpf_probe_read_kernel() invokes
> copy_from_kernel_nofault() in turn and then invokes __get_user_asm().
> Because the vsyscall page address is not readable for kernel space,
> a page fault exception is triggered accordingly, handle_page_fault()
> considers the vsyscall page address as a userspace address instead of a
> kernel space address, so the fix-up set-up by bpf isn't applied. Because
> the CPU has SMAP feature and the access happens in kernel mode, so
> page_fault_oops() is invoked and an oops happens. If these is no SMAP
> feature, the fix-up set-up by bpf will be applied and
> copy_from_kernel_nofault() will return -EFAULT instead.
> 

I find this paragraph to be a bit hard to follow. I think we can
minimize the reference to SMAP here since it is only helping detect
cross address space accesses.  How about something like the following:

The oops is triggered when:

1) A bpf program uses bpf_probe_read_kernel() to read from the vsyscall
page and invokes copy_from_kernel_nofault() which in turn calls
__get_user_asm().

2) Because the vsyscall page address is not readable from kernel space,
a page fault exception is triggered accordingly.

3) handle_page_fault() considers the vsyscall page address as a user
space address instead of a kernel space address. This results in the
fix-up setup by bpf not being applied and a page_fault_oops() is invoked
due to SMAP.

> Considering handle_page_fault() has already considered the vsyscall page
> address as a userspace address, fix the problem by disallowing vsyscall
> page read for copy_from_kernel_nofault().
> 

I agree, following the same approach as handle_page_fault() seems
reasonable.

> Originally-by: Thomas Gleixner <[e-mail]>
> Reported-by: [e-mail]
> Closes: https://lore.kernel.org/bpf/CAG48ez06TZft=[e-mail]
> Reported-by: xingwei lee <[e-mail]>
> Closes: https://lore.kernel.org/bpf/CABOYnLynjBoFZOf3Z4BhaZkc5hx_kHfsjiW+UWLoB=[e-mail]
> Signed-off-by: Hou Tao <[e-mail]>
> ---
>  arch/x86/mm/maccess.c | 9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/arch/x86/mm/maccess.c b/arch/x86/mm/maccess.c
> index 6993f026adec9..d9272e1db5224 100644
> --- a/arch/x86/mm/maccess.c
> +++ b/arch/x86/mm/maccess.c
> @@ -3,6 +3,8 @@
>  #include <linux/uaccess.h>
>  #include <linux/kernel.h>
>  
> +#include <asm/vsyscall.h>
> +
>  #ifdef CONFIG_X86_64
>  bool copy_from_kernel_nofault_allowed(const void *unsafe_src, size_t size)
>  {
> @@ -15,6 +17,13 @@ bool copy_from_kernel_nofault_allowed(const void *unsafe_src, size_t size)
>  	if (vaddr < TASK_SIZE_MAX + PAGE_SIZE)
>  		return false;
>  
> +	/* Also consider the vsyscall page as userspace address. Otherwise,
> +	 * reading the vsyscall page in copy_from_kernel_nofault() may
> +	 * trigger an oops due to an unhandled page fault.
> +	 */

x86 prefers a slightly different style for multi-line comments. Please
refer to https://docs.kernel.org/process/maintainer-tip.html#comment-style.

How about rewording the above as:

/*
 * Reading from the vsyscall page may cause an unhandled fault in
 * certain cases.  Though it is at an address above TASK_SIZE_MAX, it is
 * usually considered as a user space address.
 */


> +	if (is_vsyscall_vaddr(vaddr))
> +		return false;
> +

It would have been convenient if we had a common check for whether a
particular address is a kernel address or not. fault_in_kernel_space()
serves that purpose to an extent in other places.

I thought we could rename fault_in_kernel_space() to
vaddr_in_kernel_space() and use it here. But the check in
copy_from_kernel_nofault_allowed() includes the user guard page as well.
So the checks wouldn't exactly be the same.

I am unsure of the implications if we get rid of that difference. Maybe
we can leave it as-is for now unless someone else chimes in.

Sohil

Hou Tao Jan. 30, 2024, 4:18 a.m. UTC | #2

Hi,

On 1/30/2024 7:50 AM, Sohil Mehta wrote:
> Hi Hou Tao,
>
> I agree to your approach in this patch. Please see some comments below.
>
> On 1/26/2024 3:54 AM, Hou Tao wrote:
>> From: Hou Tao <[e-mail]>
>>
>> When trying to use copy_from_kernel_nofault() to read vsyscall page
>> through a bpf program, the following oops was reported:

[SNIP]
>> It seems the occurrence of oops depends on SMAP feature of CPU. It
>> happens as follow: a bpf program uses bpf_probe_read_kernel() to read
>> from vsyscall page, bpf_probe_read_kernel() invokes
>> copy_from_kernel_nofault() in turn and then invokes __get_user_asm().
>> Because the vsyscall page address is not readable for kernel space,
>> a page fault exception is triggered accordingly, handle_page_fault()
>> considers the vsyscall page address as a userspace address instead of a
>> kernel space address, so the fix-up set-up by bpf isn't applied. Because
>> the CPU has SMAP feature and the access happens in kernel mode, so
>> page_fault_oops() is invoked and an oops happens. If these is no SMAP
>> feature, the fix-up set-up by bpf will be applied and
>> copy_from_kernel_nofault() will return -EFAULT instead.
>>
> I find this paragraph to be a bit hard to follow. I think we can
> minimize the reference to SMAP here since it is only helping detect
> cross address space accesses.  How about something like the following:
>
> The oops is triggered when:
>
> 1) A bpf program uses bpf_probe_read_kernel() to read from the vsyscall
> page and invokes copy_from_kernel_nofault() which in turn calls
> __get_user_asm().
>
> 2) Because the vsyscall page address is not readable from kernel space,
> a page fault exception is triggered accordingly.
>
> 3) handle_page_fault() considers the vsyscall page address as a user
> space address instead of a kernel space address. This results in the
> fix-up setup by bpf not being applied and a page_fault_oops() is invoked
> due to SMAP.

Thanks for the rephrasing. It is much better now.
>> Considering handle_page_fault() has already considered the vsyscall page
>> address as a userspace address, fix the problem by disallowing vsyscall
>> page read for copy_from_kernel_nofault().
>>
> I agree, following the same approach as handle_page_fault() seems
> reasonable.
>
>> Originally-by: Thomas Gleixner <[e-mail]>
>> Reported-by: [e-mail]
>> Closes: https://lore.kernel.org/bpf/CAG48ez06TZft=[e-mail]
>> Reported-by: xingwei lee <[e-mail]>
>> Closes: https://lore.kernel.org/bpf/CABOYnLynjBoFZOf3Z4BhaZkc5hx_kHfsjiW+UWLoB=[e-mail]
>> Signed-off-by: Hou Tao <[e-mail]>
>> ---
>>  arch/x86/mm/maccess.c | 9 +++++++++
>>  1 file changed, 9 insertions(+)
>>
>> diff --git a/arch/x86/mm/maccess.c b/arch/x86/mm/maccess.c
>> index 6993f026adec9..d9272e1db5224 100644
>> --- a/arch/x86/mm/maccess.c
>> +++ b/arch/x86/mm/maccess.c
>> @@ -3,6 +3,8 @@
>>  #include <linux/uaccess.h>
>>  #include <linux/kernel.h>
>>  
>> +#include <asm/vsyscall.h>
>> +
>>  #ifdef CONFIG_X86_64
>>  bool copy_from_kernel_nofault_allowed(const void *unsafe_src, size_t size)
>>  {
>> @@ -15,6 +17,13 @@ bool copy_from_kernel_nofault_allowed(const void *unsafe_src, size_t size)
>>  	if (vaddr < TASK_SIZE_MAX + PAGE_SIZE)
>>  		return false;
>>  
>> +	/* Also consider the vsyscall page as userspace address. Otherwise,
>> +	 * reading the vsyscall page in copy_from_kernel_nofault() may
>> +	 * trigger an oops due to an unhandled page fault.
>> +	 */
> x86 prefers a slightly different style for multi-line comments. Please
> refer to https://docs.kernel.org/process/maintainer-tip.html#comment-style.

I see. Will update.
>
> How about rewording the above as:
>
> /*
>  * Reading from the vsyscall page may cause an unhandled fault in
>  * certain cases.  Though it is at an address above TASK_SIZE_MAX, it is
>  * usually considered as a user space address.
>  */

Thanks for the rewording. Will do in v3.
>
>> +	if (is_vsyscall_vaddr(vaddr))
>> +		return false;
>> +
> It would have been convenient if we had a common check for whether a
> particular address is a kernel address or not. fault_in_kernel_space()
> serves that purpose to an extent in other places.
>
> I thought we could rename fault_in_kernel_space() to
> vaddr_in_kernel_space() and use it here. But the check in
> copy_from_kernel_nofault_allowed() includes the user guard page as well.
> So the checks wouldn't exactly be the same.
>
> I am unsure of the implications if we get rid of that difference. Maybe
> we can leave it as-is for now unless someone else chimes in.

There is other difference between fault_in_kernel_space() and
copy_from_kernel_nofault_allowed(). fault_in_kernel_space() uses address
>= TASK_SIZE_MAX to check the kernel space address, but
copy_from_kernel_nofault_allowed() uses vaddr >= TASK_SIZE_MAX +
PAGE_SIZE to check the kernel space address, so I prefer to keep it as-is.
>
> Sohil

Patch

diff --git a/arch/x86/mm/maccess.c b/arch/x86/mm/maccess.c
index 6993f026adec9..d9272e1db5224 100644
--- a/arch/x86/mm/maccess.c
+++ b/arch/x86/mm/maccess.c
@@ -3,6 +3,8 @@ 
 #include <linux/uaccess.h>
 #include <linux/kernel.h>
 
+#include <asm/vsyscall.h>
+
 #ifdef CONFIG_X86_64
 bool copy_from_kernel_nofault_allowed(const void *unsafe_src, size_t size)
 {
@@ -15,6 +17,13 @@  bool copy_from_kernel_nofault_allowed(const void *unsafe_src, size_t size)
 	if (vaddr < TASK_SIZE_MAX + PAGE_SIZE)
 		return false;
 
+	/* Also consider the vsyscall page as userspace address. Otherwise,
+	 * reading the vsyscall page in copy_from_kernel_nofault() may
+	 * trigger an oops due to an unhandled page fault.
+	 */
+	if (is_vsyscall_vaddr(vaddr))
+		return false;
+
 	/*
 	 * Allow everything during early boot before 'x86_virt_bits'
 	 * is initialized.  Needed for instruction decoding in early