BPF: Fix potential bad pointer dereference in bpf_sys_bpf
PATCH 1/1 BPF: Fix potential bad pointer dereference in bpf_sys_bpf
The bpf_sys_bpf() helper function allows an eBPF program to load another eBPF program from within the kernel. In this case the argument union bpf_attr pointer (as well as the insns and license pointers inside) is a kernel address instead of a userspace address (which is the case of a usual bpf() syscall). To make the memory copying process in the syscall work in both cases, bpfptr_t [1] was introduced to wrap around the pointer and distinguish its origin. Specifically, when copying memory contents from a bpfptr_t, a copy_from_user() is performed in case of a userspace address and a memcpy() is performed for a kernel address [2].
This can lead to problems because the in-kernel pointer is never checked for validity. If an eBPF syscall program tries to call bpf_sys_bpf() with a bad insns pointer, say 0xdeadbeef (which is supposed to point to the start of the instruction array) in the bpf_attr union, memcpy() is always happy to dereference the bad pointer to cause a un-handle-able page fault and in turn an oops. However, this is not supposed to happen because at that point the eBPF program is already verified and should not cause a memory error. The same issue in userspace is handled gracefully by copy_from_user(), which would return -EFAULT in such a case.
Replace memcpy() with the safer copy_from_kernel_nofault() and strncpy_from_kernel_nofault().
[1]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/bpfptr.h
[2]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/sockptr.h#n44
--- include/linux/sockptr.h | 11 +++-------- 1 file changed, 3 insertions(+), 8 deletions(-)
include/linux/sockptr.h +3 -8
@@ -46,8 +46,7 @@ static inline int copy_from_sockptr_offset(void *dst, sockptr_t src, {if (!sockptr_is_kernel (src))return copy_from_user (dst, src.user + offset, size); -memcpy (dst, src.kernel + offset, size); -return 0 ; +return copy_from_kernel_nofault (dst, src.kernel + offset, size); }static inline int copy_from_sockptr (void *dst,sockptr_t src,size_t size)
@@ -93,12 +92,8 @@ static inline void *memdup_sockptr_nul(sockptr_t src, size_t len)static inline long strncpy_from_sockptr (char *dst,sockptr_t src,size_t count) { -if (sockptr_is_kernel (src)) { -size_t len =min (strnlen (src.kernel, count -1 ) +1 , count); - -memcpy (dst, src.kernel, len); -return len; - } +if (sockptr_is_kernel (src)) +return strncpy_from_kernel_nofault (dst, src.kernel, count);return strncpy_from_user (dst, src.user, count); }
base-commit: d295daf505758f9a0e4d05f4ee3bfdfb4192c18f
Signature · 1 line
2.35.1
Quoted 37 lines · Jinghao Jia
The
bpf_sys_bpf()helper function allows an eBPF program to load another eBPF program from within the kernel. In this case the argument unionbpf_attrpointer (as well as the insns and license pointers inside) is a kernel address instead of a userspace address (which is the case of a usualbpf()syscall). To make the memory copying process in the syscall work in both cases,bpfptr_t[1] was introduced to wrap around the pointer and distinguish its origin. Specifically, when copying memory contents from abpfptr_t, acopy_from_user()is performed in case of a userspace address and amemcpy()is performed for a kernel address [2].This can lead to problems because the in-kernel pointer is never checked for validity. If an eBPF syscall program tries to call
bpf_sys_bpf()with a bad insns pointer, say0xdeadbeef(which is supposed to point to the start of the instruction array) in thebpf_attrunion,memcpy()is always happy to dereference the bad pointer to cause a un-handle-able page fault and in turn an oops. However, this is not supposed to happen because at that point the eBPF program is already verified and should not cause a memory error. The same issue in userspace is handled gracefully bycopy_from_user(), which would return-EFAULTin such a case.Replace
memcpy()with the safercopy_from_kernel_nofault()andstrncpy_from_kernel_nofault().[1]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/bpfptr.h
[2]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/sockptr.h#n44Signed-off-by Jinghao Jia <[email protected]>--- include/linux/sockptr.h | 11 +++-------- 1 file changed, 3 insertions(+), 8 deletions(-)include/linux/sockptr.h +0
-0
include/linux/sockptr.h static inline int copy_from_sockptr_offset(void *dst, sockptr_t src, · line 50
Quoted 7 lines
include/linux/sockptr.h +1
-2@@ -46,8 +46,7 @@ static inline int copy_from_sockptr_offset(void *dst, sockptr_t src, {if (!sockptr_is_kernel (src))return copy_from_user (dst, src.user + offset, size); -memcpy (dst, src.kernel + offset, size); -return 0 ; +return copy_from_kernel_nofault (dst, src.kernel + offset, size); }
The subject and commit message mentioned it is bpf_sys_bpf() helper
might have issues. But the patch itself tries to modify copy_from_sockptr_offset() and strncpy_from_sockptr(), why?
Quoted 19 lines
include/linux/sockptr.h +2
-6static inline int copy_from_sockptr (void *dst,sockptr_t src,size_t size)@@ -93,12 +92,8 @@ static inline void *memdup_sockptr_nul(sockptr_t src, size_t len)static inline long strncpy_from_sockptr (char *dst,sockptr_t src,size_t count) { -if (sockptr_is_kernel (src)) { -size_t len =min (strnlen (src.kernel, count -1 ) +1 , count); - -memcpy (dst, src.kernel, len); -return len; - } +if (sockptr_is_kernel (src)) +return strncpy_from_kernel_nofault (dst, src.kernel, count);return strncpy_from_user (dst, src.user, count); }base-commit:
d295daf505758f9a0e4d05f4ee3bfdfb4192c18f
Quoted 35 lines · Yonghong Song
Quoted 32 lines · Jinghao Jia
The
bpf_sys_bpf()helper function allows an eBPF program to load another eBPF program from within the kernel. In this case the argument unionbpf_attrpointer (as well as the insns and license pointers inside) is a kernel address instead of a userspace address (which is the case of a usualbpf()syscall). To make the memory copying process in the syscall work in both cases,bpfptr_t[1] was introduced to wrap around the pointer and distinguish its origin. Specifically, when copying memory contents from abpfptr_t, acopy_from_user()is performed in case of a userspace address and amemcpy()is performed for a kernel address [2].This can lead to problems because the in-kernel pointer is never checked for validity. If an eBPF syscall program tries to call
bpf_sys_bpf()with a bad insns pointer, say0xdeadbeef(which is supposed to point to the start of the instruction array) in thebpf_attrunion,memcpy()is always happy to dereference the bad pointer to cause a un-handle-able page fault and in turn an oops. However, this is not supposed to happen because at that point the eBPF program is already verified and should not cause a memory error. The same issue in userspace is handled gracefully bycopy_from_user(), which would return-EFAULTin such a case.Replace
memcpy()with the safercopy_from_kernel_nofault()andstrncpy_from_kernel_nofault().[1]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/bpfptr.h
[2]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/sockptr.h#n44Signed-off-by Jinghao Jia <[email protected]>---
include/linux/sockptr.h| 11 +++--------1 file changed, 3 insertions(+), 8 deletions(-)
Quoted 17 lines
include/linux/sockptr.h static inline int copy_from_sockptr_offset(void · line 51
Quoted 13 lines
include/linux/sockptr.h +1
-2@@ -46,8 +46,7 @@ static inline int copy_from_sockptr_offset(void *dst, sockptr_t src, { if (!sockptr_is_kernel(src)) return copy_from_user(dst, src.user + offset, size); - memcpy(dst, src.kernel + offset, size); - return 0; + return copy_from_kernel_nofault(dst, src.kernel + offset, size); }The subject and commit message mentioned it is
bpf_sys_bpf()helper
might have issues. But the patch itself tries to modifycopy_from_sockptr_offset()andstrncpy_from_sockptr(), why?
Hi Yonghong,
Sorry for the confusion. The problem happens when bpf_sys_bpf() helper is called with a bad kernel address but the dereference takes place in the copy_from_sockptr_offset() and strncpy_from_sockptr() functions.
Let's assume we are doing a BPF_PROG_LOAD operation using bpf_sys_bpf() and our insns pointer inside the union bpf_attr argument is set to NULL (could be any other bad address). The helper calls __sys_bpf() which would then call bpf_prog_load() to load the program. bpf_prog_load() is responsible for copying the eBPF instructions to the newly allocated memory for the program, which uses the bpfptr_t API [1]. Internally, all bpfptr_t operations are backed by the corresponding sockptr_t operations. In other words, the code that performs the copy (and therefore the deref of the pointer) is inside copy_from_sockptr_offset() and strncpy_from_sockptr().
[1]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/kernel/bpf/syscall.c#n2566
Best,
Jinghao
Quoted 17 lines
Quoted 17 lines
include/linux/sockptr.h +2
-6static inline int copy_from_sockptr(void *dst, sockptr_t src, size_t size)@@ -93,12 +92,8 @@ static inline void *memdup_sockptr_nul(sockptr_t src, size_t len) static inline long strncpy_from_sockptr(char *dst, sockptr_t src, size_t count) { - if (sockptr_is_kernel(src)) { - size_t len = min(strnlen(src.kernel, count - 1) + 1, count); - - memcpy(dst, src.kernel, len); - return len; - } + if (sockptr_is_kernel(src)) + return strncpy_from_kernel_nofault(dst, src.kernel, count); return strncpy_from_user(dst, src.user, count); }base-commit:
d295daf505758f9a0e4d05f4ee3bfdfb4192c18f
Quoted 57 lines · Jinghao Jia
Quoted 37 lines · Yonghong Song
Quoted 34 lines · Jinghao Jia
The
bpf_sys_bpf()helper function allows an eBPF program to load another eBPF program from within the kernel. In this case the argument unionbpf_attrpointer (as well as the insns and license pointers inside) is a kernel address instead of a userspace address (which is the case of a usualbpf()syscall). To make the memory copying process in the syscall work in both cases,bpfptr_t[1] was introduced to wrap around the pointer and distinguish its origin. Specifically, when copying memory contents from abpfptr_t, acopy_from_user()is performed in case of a userspace address and amemcpy()is performed for a kernel address [2].This can lead to problems because the in-kernel pointer is never checked for validity. If an eBPF syscall program tries to call
bpf_sys_bpf()with a bad insns pointer, say0xdeadbeef(which is supposed to point to the start of the instruction array) in thebpf_attrunion,memcpy()is always happy to dereference the bad pointer to cause a un-handle-able page fault and in turn an oops. However, this is not supposed to happen because at that point the eBPF program is already verified and should not cause a memory error. The same issue in userspace is handled gracefully bycopy_from_user(), which would return-EFAULTin such a case.Replace
memcpy()with the safercopy_from_kernel_nofault()andstrncpy_from_kernel_nofault().[1]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/bpfptr.h
[2]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/sockptr.h#n44
Signed-off-by Jinghao Jia <[email protected]>---
include/linux/sockptr.h| 11 +++--------1 file changed, 3 insertions(+), 8 deletions(-)Quoted 17 lines
include/linux/sockptr.h static inline int copy_from_sockptr_offset(void · line 51
Quoted 13 lines
include/linux/sockptr.h +1
-2@@ -46,8 +46,7 @@ static inline int copy_from_sockptr_offset(void *dst, sockptr_t src, { if (!sockptr_is_kernel(src)) return copy_from_user(dst, src.user + offset, size); - memcpy(dst, src.kernel + offset, size); - return 0; + return copy_from_kernel_nofault(dst, src.kernel + offset, size); }The subject and commit message mentioned it is
bpf_sys_bpf()helper
might have issues. But the patch itself tries to modifycopy_from_sockptr_offset()andstrncpy_from_sockptr(), why?
Quoted 5 lines
Hi Yonghong,
Sorry for the confusion. The problem happens when
bpf_sys_bpf()helper is called with a bad kernel address but the dereference takes place in thecopy_from_sockptr_offset()andstrncpy_from_sockptr()functions.Let's assume we are doing a
BPF_PROG_LOADoperation usingbpf_sys_bpf()and our insns pointer inside the unionbpf_attrargument is set to NULL (could be any other bad address). The helper calls__sys_bpf()which would then callbpf_prog_load()to load the program.bpf_prog_load()is responsible for copying the eBPF instructions to the newly allocated memory for the program, which uses thebpfptr_tAPI [1]. Internally, allbpfptr_toperations are backed by the correspondingsockptr_toperations. In other words, the code that performs the copy (and therefore the deref of the pointer) is insidecopy_from_sockptr_offset()andstrncpy_from_sockptr().
Thanks for explanation. It would be great if you can put the above details in the commit message (esp. call stack) which leads to the kernel panic(?).
Quoted 6 lines
[1]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/kernel/bpf/syscall.c#n2566
Best,
Jinghao
Quoted 15 lines
Quoted 7 lines
Quoted 7 lines
include/linux/sockptr.h +0
-2static inline int copy_from_sockptr(void *dst, sockptr_t src, size_t size)@@ -93,12 +92,8 @@ static inline void *memdup_sockptr_nul(sockptr_t src, size_t len) static inline long strncpy_from_sockptr(char *dst, sockptr_t src, size_t count) { - if (sockptr_is_kernel(src)) { - size_t len = min(strnlen(src.kernel, count - 1) + 1, count);Quoted 8 lines
include/linux/sockptr.h static inline void *memdup_sockptr_nul(sockptr_t src, size_t len) · line 98
Quoted 8 lines
include/linux/sockptr.h +2
-4- - memcpy(dst, src.kernel, len); - return len; - } + if (sockptr_is_kernel(src)) + return strncpy_from_kernel_nofault(dst, src.kernel, count); return strncpy_from_user(dst, src.user, count); }
I think we should not modify copy_from_sockptr() and strncpy_from_sockptr(). These two functions are used by networking as well and nofault version should not be called for calls in
networking stack.
So I suggest you directly modify copy_from_bpfptr() and strncpy_from_bpfptr() since these two functions indeed might have invalid kernel address and may cause fault.
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
base-commit:
d295daf505758f9a0e4d05f4ee3bfdfb4192c18f
Quoted 96 lines · Yonghong Song
Quoted 57 lines · Jinghao Jia
Quoted 37 lines · Yonghong Song
Quoted 34 lines · Jinghao Jia
The
bpf_sys_bpf()helper function allows an eBPF program to load another eBPF program from within the kernel. In this case the argument unionbpf_attrpointer (as well as the insns and license pointers inside) is a kernel address instead of a userspace address (which is the case of a usualbpf()syscall). To make the memory copying process in the syscall work in both cases,bpfptr_t[1] was introduced to wrap around the pointer and distinguish its origin. Specifically, when copying memory contents from abpfptr_t, acopy_from_user()is performed in case of a userspace address and amemcpy()is performed for a kernel address [2].This can lead to problems because the in-kernel pointer is never checked for validity. If an eBPF syscall program tries to call
bpf_sys_bpf()with a bad insns pointer, say0xdeadbeef(which is supposed to point to the start of the instruction array) in thebpf_attrunion,memcpy()is always happy to dereference the bad pointer to cause a un-handle-able page fault and in turn an oops. However, this is not supposed to happen because at that point the eBPF program is already verified and should not cause a memory error. The same issue in userspace is handled gracefully bycopy_from_user(), which would return-EFAULTin such a case.Replace
memcpy()with the safercopy_from_kernel_nofault()andstrncpy_from_kernel_nofault().[1]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/bpfptr.h
[2]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/include/linux/sockptr.h#n44
Signed-off-by Jinghao Jia <[email protected]>---
include/linux/sockptr.h| 11 +++--------1 file changed, 3 insertions(+), 8 deletions(-)Quoted 17 lines
include/linux/sockptr.h static inline int copy_from_sockptr_offset(void · line 51
Quoted 13 lines
include/linux/sockptr.h +1
-2@@ -46,8 +46,7 @@ static inline int copy_from_sockptr_offset(void *dst, sockptr_t src, { if (!sockptr_is_kernel(src)) return copy_from_user(dst, src.user + offset, size); - memcpy(dst, src.kernel + offset, size); - return 0; + return copy_from_kernel_nofault(dst, src.kernel + offset, size); }The subject and commit message mentioned it is
bpf_sys_bpf()helper
might have issues. But the patch itself tries to modifycopy_from_sockptr_offset()andstrncpy_from_sockptr(), why?Quoted 5 lines
Hi Yonghong,
Sorry for the confusion. The problem happens when
bpf_sys_bpf()helper is called with a bad kernel address but the dereference takes place in thecopy_from_sockptr_offset()andstrncpy_from_sockptr()functions.Let's assume we are doing a
BPF_PROG_LOADoperation usingbpf_sys_bpf()and our insns pointer inside the unionbpf_attrargument is set to NULL (could be any other bad address). The helper calls__sys_bpf()which would then callbpf_prog_load()to load the program.bpf_prog_load()is responsible for copying the eBPF instructions to the newly allocated memory for the program, which uses thebpfptr_tAPI [1]. Internally, allbpfptr_toperations are backed by the correspondingsockptr_toperations. In other words, the code that performs the copy (and therefore the deref of the pointer) is insidecopy_from_sockptr_offset()andstrncpy_from_sockptr().Thanks for explanation. It would be great if you can put the above details in the commit message (esp. call stack) which leads to the kernel panic(?).
Quoted 6 lines
[1]: https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/kernel/bpf/syscall.c#n2566
Best,
JinghaoQuoted 15 lines
Quoted 7 lines
Quoted 7 lines
include/linux/sockptr.h +0
-2static inline int copy_from_sockptr(void *dst, sockptr_t src, size_t size)@@ -93,12 +92,8 @@ static inline void *memdup_sockptr_nul(sockptr_t src, size_t len) static inline long strncpy_from_sockptr(char *dst, sockptr_t src, size_t count) { - if (sockptr_is_kernel(src)) { - size_t len = min(strnlen(src.kernel, count - 1) + 1, count);Quoted 8 lines
include/linux/sockptr.h static inline void *memdup_sockptr_nul(sockptr_t src, size_t len) · line 98
Quoted 8 lines
include/linux/sockptr.h +2
-4- - memcpy(dst, src.kernel, len); - return len; - } + if (sockptr_is_kernel(src)) + return strncpy_from_kernel_nofault(dst, src.kernel, count); return strncpy_from_user(dst, src.user, count); }I think we should not modify
copy_from_sockptr()andstrncpy_from_sockptr(). These two functions are used by networking as well and nofault version should not be called for calls in
networking stack.
Quoted 3 lines
So I suggest you directly modify
copy_from_bpfptr()andstrncpy_from_bpfptr()since these two functions indeed might have invalid kernel address and may cause fault.
Thanks a lot for the feedback! I do agree that the changes on sockptr functions might be problematic. I will roll out a V2 based on your comments.
--Jinghao
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
base-commit:
d295daf505758f9a0e4d05f4ee3bfdfb4192c18f