On Fri, Sep 27, 2024 at 06:52:09PM GMT, jeffxu@xxxxxxxxxxxx wrote:
> From: Jeff Xu <jeffxu@xxxxxxxxxxxx>
>
> Update doc after in-loop change: mprotect/madvise can have
> partially updated and munmap is atomic.
>
> Fix indentation and clarify some sections to improve readability.
>
> Signed-off-by: Jeff Xu <jeffxu@xxxxxxxxxxxx>
> ---
> Documentation/userspace-api/mseal.rst | 290 ++++++++++++--------------
> 1 file changed, 136 insertions(+), 154 deletions(-)
>
> diff --git a/Documentation/userspace-api/mseal.rst b/Documentation/userspace-api/mseal.rst
> index 4132eec995a3..68986084e191 100644
> --- a/Documentation/userspace-api/mseal.rst
> +++ b/Documentation/userspace-api/mseal.rst
> @@ -23,177 +23,159 @@ applications can additionally seal security critical data at runtime.
> A similar feature already exists in the XNU kernel with the
> VM_FLAGS_PERMANENT flag [1] and on OpenBSD with the mimmutable syscall [2].
>
> -User API
> -========
> -mseal()
> ------------
> -The mseal() syscall has the following signature:
> -
> -``int mseal(void addr, size_t len, unsigned long flags)``
> -
> -**addr/len**: virtual memory address range.
> -
> -The address range set by ``addr``/``len`` must meet:
> - - The start address must be in an allocated VMA.
> - - The start address must be page aligned.
> - - The end address (``addr`` + ``len``) must be in an allocated VMA.
> - - no gap (unallocated memory) between start and end address.
> -
> -The ``len`` will be paged aligned implicitly by the kernel.
> -
> -**flags**: reserved for future use.
> -
> -**return values**:
> -
> -- ``0``: Success.
> -
> -- ``-EINVAL``:
> - - Invalid input ``flags``.
> - - The start address (``addr``) is not page aligned.
> - - Address range (``addr`` + ``len``) overflow.
> -
> -- ``-ENOMEM``:
> - - The start address (``addr``) is not allocated.
> - - The end address (``addr`` + ``len``) is not allocated.
> - - A gap (unallocated memory) between start and end address.
> -
> -- ``-EPERM``:
> - - sealing is supported only on 64-bit CPUs, 32-bit is not supported.
> -
> -- For above error cases, users can expect the given memory range is
> - unmodified, i.e. no partial update.
> -
> -- There might be other internal errors/cases not listed here, e.g.
> - error during merging/splitting VMAs, or the process reaching the max
> - number of supported VMAs. In those cases, partial updates to the given
> - memory range could happen. However, those cases should be rare.
> -
> -**Blocked operations after sealing**:
> - Unmapping, moving to another location, and shrinking the size,
> - via munmap() and mremap(), can leave an empty space, therefore
> - can be replaced with a VMA with a new set of attributes.
> -
> - Moving or expanding a different VMA into the current location,
> - via mremap().
> -
> - Modifying a VMA via mmap(MAP_FIXED).
> -
> - Size expansion, via mremap(), does not appear to pose any
> - specific risks to sealed VMAs. It is included anyway because
> - the use case is unclear. In any case, users can rely on
> - merging to expand a sealed VMA.
> -
> - mprotect() and pkey_mprotect().
> -
> - Some destructive madvice() behaviors (e.g. MADV_DONTNEED)
> - for anonymous memory, when users don't have write permission to the
> - memory. Those behaviors can alter region contents by discarding pages,
> - effectively a memset(0) for anonymous memory.
> -
> - Kernel will return -EPERM for blocked operations.
> -
> - For blocked operations, one can expect the given address is unmodified,
> - i.e. no partial update. Note, this is different from existing mm
> - system call behaviors, where partial updates are made till an error is
> - found and returned to userspace. To give an example:
> -
> - Assume following code sequence:
> -
> - - ptr = mmap(null, 8192, PROT_NONE);
> - - munmap(ptr + 4096, 4096);
> - - ret1 = mprotect(ptr, 8192, PROT_READ);
> - - mseal(ptr, 4096);
> - - ret2 = mprotect(ptr, 8192, PROT_NONE);
> -
> - ret1 will be -ENOMEM, the page from ptr is updated to PROT_READ.
> -
> - ret2 will be -EPERM, the page remains to be PROT_READ.
> -
> -**Note**:
> -
> -- mseal() only works on 64-bit CPUs, not 32-bit CPU.
> -
> -- users can call mseal() multiple times, mseal() on an already sealed memory
> - is a no-action (not error).
> -
> -- munseal() is not supported.
> +SYSCALL
> +=======
> +mseal syscall signature
> +-----------------------
> + **int** mseal(**void \*** addr, **size_t** len, **unsigned long** flags)
> +
> + **addr**/**len**: virtual memory address range.
> + The address range set by **addr**/**len** must meet:
> + - The start address must be in an allocated VMA.
> + - The start address must be page aligned.
> + - The end address (**addr** + **len**) must be in an allocated VMA.
> + - no gap (unallocated memory) between start and end address.
> +
> + The ``len`` will be paged aligned implicitly by the kernel.
> +
> + **flags**: reserved for future use.
> +
> + **Return values**:
> + - **0**: Success.
> + - **-EINVAL**:
> + * Invalid input ``flags``.
> + * The start address (``addr``) is not page aligned.
> + * Address range (``addr`` + ``len``) overflow.
> + - **-ENOMEM**:
> + * The start address (``addr``) is not allocated.
> + * The end address (``addr`` + ``len``) is not allocated.
> + * A gap (unallocated memory) between start and end address.
> + - **-EPERM**:
> + * sealing is supported only on 64-bit CPUs, 32-bit is not supported.
> +
> + **Note about error return**:
> + - For above error cases, users can expect the given memory range is
> + unmodified, i.e. no partial update.
> + - There might be other internal errors/cases not listed here, e.g.
> + error during merging/splitting VMAs, or the process reaching the max
> + number of supported VMAs. In those cases, partial updates to the given
> + memory range could happen. However, those cases should be rare.
How about turning the above into a man page?
> + **Architecture support**:
> + mseal only works on 64-bit CPUs, not 32-bit CPU.
> +
> + **Idempotent**:
> + users can call mseal multiple times, mseal on an already sealed memory
> + is a no-action (not error).
> +
> + **no munseal**
> + Once mapping is sealed, it can't be unsealed. kernel should never
> + have munseal, this is consistent with other sealing feature, e.g.
> + F_SEAL_SEAL for file.
> +
> +Blocked mm syscall for sealed mapping
> +-------------------------------------
> + It might be imporant to note: **once the mapping is sealed, it will
> + stay in the process's memory till the process terminates**.
> +
> + Example::
> +
> + *ptr = mmap(0, 4096, PROT_READ, MAP_ANONYMOUS | MAP_PRIVATE, 0, 0);
> + rc = mseal(ptr, 4096, 0);
> + /* munmap will fail */
> + rc = munmap(ptr, 4096);
> + assert(rc < 0);
> +
> + Blocked mm syscall:
> + - munmap
> + - mmap
> + - mremap
> + - mprotect and pkey_mprotect
> + - some destructive madvise behaviors: MADV_DONTNEED, MADV_FREE,
> + MADV_DONTNEED_LOCKED, MADV_FREE, MADV_DONTFORK, MADV_WIPEONFORK
> +
> + The first set of syscall to block is munmap, mremap, mmap. They can
> + either leave an empty space in the address space, therefore allow
> + replacement with a new mapping with new set of attributes, or can
> + overwrite the existing mapping with another mapping.
> +
> + mprotect and pkey_mprotect are blocked because they changes the
change
> + protection bits (rwx) of the mapping.
> +
> + Some destructive madvice behaviors (MADV_DONTNEED, MADV_FREE,
> + MADV_DONTNEED_LOCKED, MADV_FREE, MADV_DONTFORK, MADV_WIPEONFORK)
> + for anonymous memory, when users don't have write permission to the
> + memory. Those behaviors can alter region contents by discarding pages,
> + effectively a memset(0) for anonymous memory.
What's the difference between anonymous memory and MAP_PRIVATE | MAP_FILE?
The feature now, as is (as far as I understand!) will allow you to do things like MADV_DONTNEED
on a read-only file mapping. e.g .text. This is obviously wrong?
> +
> + Kernel will return -EPERM for blocked syscalls.
> +
> + When blocked syscall return -EPERM due to sealing, the memory regions may or may not be changed, depends on the syscall being blocked:
> + - munmap: munmap is atomic. If one of VMAs in the given range is
> + sealed, none of VMAs are updated.
> + - mprotect, pkey_mprotect, madvise: partial update might happen, e.g.
> + when mprotect over multiple VMAs, mprotect might update the beginning
> + VMAs before reaching the sealed VMA and return -EPERM.
> + - mmap and mremap: undefined behavior.
mmap and mremap are actually not undefined as they use munmap semantics for their unmapping.
Whether this is something we'd want to document, I don't know honestly (nor do I think is ever written down in POSIX?)
>
> Use cases:
> ==========
> - glibc:
> The dynamic linker, during loading ELF executables, can apply sealing to
> - non-writable memory segments.
> + mapping segments.
>
> - Chrome browser: protect some security sensitive data-structures.
>
> -Notes on which memory to seal:
> -==============================
> -
> -It might be important to note that sealing changes the lifetime of a mapping,
> -i.e. the sealed mapping won’t be unmapped till the process terminates or the
> -exec system call is invoked. Applications can apply sealing to any virtual
> -memory region from userspace, but it is crucial to thoroughly analyze the
> -mapping's lifetime prior to apply the sealing.
> +Don't use mseal on:
> +===================
> +Applications can apply sealing to any virtual memory region from userspace,
> +but it is *crucial to thoroughly analyze the mapping's lifetime* prior to
> +apply the sealing. This is because the sealed mapping *won’t be unmapped*
> +till the process terminates or the exec system call is invoked.
There should probably be a nice disclaimer as to how most people don't need this or shouldn't use this.
At least in its current form.
<snip>
> -
> -
> -Additional notes:
> -=================
> As Jann Horn pointed out in [3], there are still a few ways to write
> -to RO memory, which is, in a way, by design. Those cases are not covered
> -by mseal(). If applications want to block such cases, sandbox tools (such as
> -seccomp, LSM, etc) might be considered.
> +to RO memory, which is, in a way, by design. And those could be blocked
> +by different security measures.
>
> Those cases are:
> -
> -- Write to read-only memory through /proc/self/mem interface.
> -- Write to read-only memory through ptrace (such as PTRACE_POKETEXT).
> -- userfaultfd.
> + - Write to read-only memory through /proc/self/mem interface (FOLL_FORCE).
> + - Write to read-only memory through ptrace (such as PTRACE_POKETEXT).
> + - userfaultfd.
I don't understand how this is not a problem, but MADV_DONTNEED is.
To me it seems that what we have now is completely useless, because you can trivially
bypass it using /proc/self/mem, which is enabled on most Linux systems.
Before you mention ChromeOS or Chrome, I don't care. Kernel features aren't designed
for Chrome. They need to work with every other distro and application as well.
It seems to me that the most sensible change is blocking/somehow distinguishing between /proc/self/mem and
/proc/<pid>/mem (some other process) and ptrace. As in blocking /proc/self/mem but allowing the other FOLL_FORCE's
as the traditional UNIX permission model allows.
--
Pedro
