Skip to contentSkip to conversation

mm/secretmem: memory address mapped to memfd_secret can be used in write syscall.

BUG? David Wang 3 participants 5 messages Nov 8, 2023 – Nov 13, 2023 Memory management

lore
Share & tools
lore

David Wang

author
#0 lore raw

Hi,
According to https://lwn.net/Articles/865256/,
the memory address got from memfd_secret/ftruncate/mmap should not be used by syscalls, since it is not accessible even by kernel.

But my test result shows that the "secret" memory could be used in syscall write, is this expected behavior? This is my test code:

int main() {
	int fd = syscall(__NR_memfd_secret, 0);
	if (fd < 0) {
		perror("Fail to create secret");
		return -1;
	}
	if (ftruncate(fd, 1024) < 0) {
		perror("Fail to size the secret");
		return -1;
	}
	char *key = mmap(NULL, 1024, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
	if (key == MAP_FAILED) {
		perror("Fail to mmap");
		return -1;
	}
	// should be some secure channel
	strcpy(key, "ThisIsAKey");
	// printf("[%d]key(%s) ready: %p\n", getpid(), key, key);
	// getchar();
	// make syscall, should err
	write(STDOUT_FILENO, key, strlen(key));  //<-- Here the key shows up on stdout.

	return 0;
}

Thanks
David

#1 lore raw
Quoted 7 lines · David Wang

Hi,
According to https://lwn.net/Articles/865256/,
the memory address got from memfd_secret/ftruncate/mmap should not be used by syscalls, since it is not accessible even by kernel.

But my test result shows that the "secret" memory could be used in syscall write, is this expected behavior? This is my test code:

CCing Mike.

According to the man page:

"The memory areas backing the file created with memfd_secret(2) are visible only to the processes that have access to the file descriptor. The memory region is removed from the kernel page tables and only the page tables of the processes holding the file descriptor map the corresponding physical memory. (Thus, the pages in the region can't be accessed by the kernel itself, so that, for example, pointers to the region can't be passed to system calls.)

I'm not sure if the last part is actually true, if the syscalls end up walking user page tables to copy data in/out.

Quoted 14 lines
int main() {
	int fd = syscall(__NR_memfd_secret, 0);
	if (fd < 0) {
		perror("Fail to create secret");
		return -1;
	}
	if (ftruncate(fd, 1024) < 0) {
		perror("Fail to size the secret");
		return -1;
	}
	char *key = mmap(NULL, 1024, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
	if (key == MAP_FAILED) {
		perror("Fail to mmap");
Quoted 8 lines
		return -1;
	}
	// should be some secure channel
	strcpy(key, "ThisIsAKey");
	// printf("[%d]key(%s) ready: %p\n", getpid(), key, key);
	// getchar();
	// make syscall, should err
	write(STDOUT_FILENO, key, strlen(key));  //<-- Here the key shows up on stdout.

What probably happens here is that the kernel reads the data via the user page tables, and can, therefore, access that memory just fine.

Looking at the selftest (tools/testing/selftests/mm/memfd_secret.c) we test that we cannot read from the memfd and cannot write to the memfd. We don't test if other syscalls can access that user-provided buffer that is backed by a memfd.

Signature · 3 lines
Cheers,

David / dhildenb
#2 lore raw
Quoted 11 lines · David Hildenbrand

According to the man page:

"The memory areas backing the file created with memfd_secret(2) are visible only to the processes that have access to the file descriptor. The memory region is removed from the kernel page tables and only the page tables of the processes holding the file descriptor map the corresponding physical memory. (Thus, the pages in the region can't be accessed by the kernel itself, so that, for example, pointers to the region can't be passed to system calls.)

Quoted 2 lines

I'm not sure if the last part is actually true, if the syscalls end up walking user page tables to copy data in/out.

The idea behind removing it from the kernel page tables is so that kernel code running in some other process context won't be able to reference the memory via the kernel address space. (So if there is some kind of kernel zero-day which allows arbitrary code execution, the injected attack code would have to play games with page tables before being able to reference the memory --- this is not *impossible*, just more annoying.)

But if you are doing a buffered write, the copy from the user-supplied buffer to the page cache is happening in the process's context. So "foreground kernel code" can dereference the user-supplied pointer just fine.

Cheers,

- Ted

#3 lore raw
Quoted 23 lines · Theodore Ts'o
Quoted 11 lines · David Hildenbrand

According to the man page:

"The memory areas backing the file created with memfd_secret(2) are visible only to the processes that have access to the file descriptor. The memory region is removed from the kernel page tables and only the page tables of the processes holding the file descriptor map the corresponding physical memory. (Thus, the pages in the region can't be accessed by the kernel itself, so that, for example, pointers to the region can't be passed to system calls.)

Quoted 2 lines

I'm not sure if the last part is actually true, if the syscalls end up walking user page tables to copy data in/out.

The idea behind removing it from the kernel page tables is so that kernel code running in some other process context won't be able to reference the memory via the kernel address space. (So if there is some kind of kernel zero-day which allows arbitrary code execution, the injected attack code would have to play games with page tables before being able to reference the memory --- this is not *impossible*, just more annoying.)

Quoted 4 lines

But if you are doing a buffered write, the copy from the user-supplied buffer to the page cache is happening in the process's context. So "foreground kernel code" can dereference the user-supplied pointer just fine.

Right, so the statement in the man page is imprecise.

Signature · 3 lines
Cheers,

David / dhildenb
author
#4 lore raw
In reply to Theodore Ts'o
Quoted 23 lines · Theodore Ts'o
Quoted 11 lines · David Hildenbrand

According to the man page:

"The memory areas backing the file created with memfd_secret(2) are visible only to the processes that have access to the file descriptor. The memory region is removed from the kernel page tables and only the page tables of the processes holding the file descriptor map the corresponding physical memory. (Thus, the pages in the region can't be accessed by the kernel itself, so that, for example, pointers to the region can't be passed to system calls.)

Quoted 2 lines

I'm not sure if the last part is actually true, if the syscalls end up walking user page tables to copy data in/out.

The idea behind removing it from the kernel page tables is so that kernel code running in some other process context won't be able to reference the memory via the kernel address space. (So if there is some kind of kernel zero-day which allows arbitrary code execution, the injected attack code would have to play games with page tables before being able to reference the memory --- this is not *impossible*, just more annoying.)

Quoted 5 lines

But if you are doing a buffered write, the copy from the user-supplied buffer to the page cache is happening in the process's context. So "foreground kernel code" can dereference the user-supplied pointer just fine.

But the inconsistent treatment in kernel, memfd denied while mmaped-address allowed, is kind of confusing...
I thought those two should be treated the same way....

Thanks
David Wang