Re: [patch 4/4] prctl: PR_SET_MM -- Introduce PR_SET_MM_MAP operation, v3

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 04 Aug 2014 21:22:59 +0400 Cyrill Gorcunov <gorcunov@openvz.org> wrote:

> During development of c/r we've noticed that in case if we need to
> support user namespaces we face a problem with capabilities in
> prctl(PR_SET_MM, ...) call, in particular once new user namespace
> is created capable(CAP_SYS_RESOURCE) no longer passes.
> 
> A approach is to eliminate CAP_SYS_RESOURCE check but pass all
> new values in one bundle, which would allow the kernel to make
> more intensive test for sanity of values and same time allow us to
> support checkpoint/restore of user namespaces.
> 
> Thus a new command PR_SET_MM_MAP introduced. It takes a pointer of
> prctl_mm_map structure which carries all the members to be updated.
> 
> 	prctl(PR_SET_MM, PR_SET_MM_MAP, struct prctl_mm_map *, size)
> 
> 	struct prctl_mm_map {
> 		__u64	start_code;
> 		__u64	end_code;
> 		__u64	start_data;
> 		__u64	end_data;
> 		__u64	start_brk;
> 		__u64	brk;
> 		__u64	start_stack;
> 		__u64	arg_start;
> 		__u64	arg_end;
> 		__u64	env_start;
> 		__u64	env_end;
> 		__u64	*auxv;
> 		__u32	auxv_size;
> 		__u32	exe_fd;
> 	};
> 
> All members except @exe_fd correspond ones of struct mm_struct.
> To figure out which available values these members may take here
> are meanings of the members.
> 
>  - start_code, end_code: represent bounds of executable code area
>  - start_data, end_data: represent bounds of data area
>  - start_brk, brk: used to calculate bounds for brk() syscall
>  - start_stack: used when accounting space needed for command
>    line arguments, environment and shmat() syscall
>  - arg_start, arg_end, env_start, env_end: represent memory area
>    supplied for command line arguments and environment variables
>  - auxv, auxv_size: carries auxiliary vector, Elf format specifics
>  - exe_fd: file descriptor number for executable link (/proc/self/exe)
> 
> Thus we apply the following requirements to the values
> 
> 1) Any member except @auxv, @auxv_size, @exe_fd is rather an address
>    in user space thus it must be laying inside [mmap_min_addr, mmap_max_addr)
>    interval.
> 
> 2) While @[start|end]_code and @[start|end]_data may point to an nonexisting
>    VMAs (say a program maps own new .text and .data segments during execution)
>    the rest of members should belong to VMA which must exist.
> 
> 3) Addresses must be ordered, ie @start_ member must not be greater or
>    equal to appropriate @end_ member.
> 
> 4) As in regular Elf loading procedure we require that @start_brk and
>    @brk be greater than @end_data.
> 
> 5) If RLIMIT_DATA rlimit is set to non-infinity new values should not
>    exceed existing limit. Same applies to RLIMIT_STACK.
> 
> 6) Auxiliary vector size must not exceed existing one (which is
>    predefined as AT_VECTOR_SIZE and depends on architecture).
> 
> 7) File descriptor passed in @exe_file should be pointing
>    to executable file (because we use existing prctl_set_mm_exe_file_locked
>    helper it ensures that the file we are going to use as exe link has all
>    required permission granted).
> 
> Now about where these members are involved inside kernel code:
> 
>  - @start_code and @end_code are used in /proc/$pid/[stat|statm] output;
> 
>  - @start_data and @end_data are used in /proc/$pid/[stat|statm] output,
>    also they are considered if there enough space for brk() syscall
>    result if RLIMIT_DATA is set;
> 
>  - @start_brk shown in /proc/$pid/stat output and accounted in brk()
>    syscall if RLIMIT_DATA is set; also this member is tested to
>    find a symbolic name of mmap event for perf system (we choose
>    if event is generated for "heap" area); one more aplication is
>    selinux -- we test if a process has PROCESS__EXECHEAP permission
>    if trying to make heap area being executable with mprotect() syscall;
> 
>  - @brk is a current value for brk() syscall which lays inside heap
>    area, it's shown in /proc/$pid/stat. When syscall brk() succesfully
>    provides new memory area to a user space upon brk() completion the
>    mm::brk is updated to carry new value;
> 
>    Both @start_brk and @brk are actively used in /proc/$pid/maps
>    and /proc/$pid/smaps output to find a symbolic name "heap" for
>    VMA being scanned;
> 
>  - @start_stack is printed out in /proc/$pid/stat and used to
>    find a symbolic name "stack" for task and threads in
>    /proc/$pid/maps and /proc/$pid/smaps output, and as the same
>    as with @start_brk -- perf system uses it for event naming.
>    Also kernel treat this member as a start address of where
>    to map vDSO pages and to check if there is enough space
>    for shmat() syscall;
> 
>  - @arg_start, @arg_end, @env_start and @env_end are printed out
>    in /proc/$pid/stat. Another access to the data these members
>    represent is to read /proc/$pid/environ or /proc/$pid/cmdline.
>    Any attempt to read these areas kernel tests with access_process_vm
>    helper so a user must have enough rights for this action;
> 
>  - @auxv and @auxv_size may be read from /proc/$pid/auxv. Strictly
>    speaking kernel doesn't care much about which exactly data is
>    sitting there because it is solely for userspace;
> 
>  - @exe_fd is referred from /proc/$pid/exe and when generating
>    coredump. We uses prctl_set_mm_exe_file_locked helper to update
>    this member, so exe-file link modification remains one-shot
>    action.
> 
> Still note that updating exe-file link now doesn't require sys-resource
> capability anymore, after all there is no much profit in preventing setup
> own file link (there are a number of ways to execute own code -- ptrace,
> ld-preload, so that the only reliable way to find which exactly code
> is executed is to inspect running program memory). Still we require
> the caller to be at least user-namespace root user.
> 
> I believe the old interface should be deprecated and ripped off
> in a couple of kernel releases if no one against.
> 
> To test if new interface is implemented in the kernel one
> can pass PR_SET_MM_MAP_SIZE opcode and the kernel returns
> the size of currently supported struct prctl_mm_map.

Please convince me that we're not adding any security holes.


> +#ifdef CONFIG_CHECKPOINT_RESTORE
> +/*
> + * WARNING: we don't require any capability here so be very careful
> + * in what is allowed for modification from userspace.
> + */
> +static int validate_prctl_map_locked(struct prctl_mm_map *prctl_map)
> +{
> +	unsigned long mmap_max_addr = TASK_SIZE;
> +	struct mm_struct *mm = current->mm;
> +	struct vm_area_struct *stack_vma;
> +	int error = 0;
> +
> +	/*
> +	 * Make sure the members are not somewhere outside
> +	 * of allowed address space.
> +	 */
> +#define __prctl_check_addr_space(__member)					\
> +	({									\
> +		int __rc;							\
> +		if ((unsigned long)prctl_map->__member < mmap_max_addr &&	\
> +		    (unsigned long)prctl_map->__member >= mmap_min_addr)	\
> +			__rc = 0;						\
> +		else								\
> +			__rc = -EINVAL;						\
> +		__rc;								\
> +	})
> +	error |= __prctl_check_addr_space(start_code);
> +	error |= __prctl_check_addr_space(end_code);
> +	error |= __prctl_check_addr_space(start_data);
> +	error |= __prctl_check_addr_space(end_data);
> +	error |= __prctl_check_addr_space(start_stack);
> +	error |= __prctl_check_addr_space(start_brk);
> +	error |= __prctl_check_addr_space(brk);
> +	error |= __prctl_check_addr_space(arg_start);
> +	error |= __prctl_check_addr_space(arg_end);
> +	error |= __prctl_check_addr_space(env_start);
> +	error |= __prctl_check_addr_space(env_end);

Boy this is verbose.  I had a little fiddle and came up with

--- a/kernel/sys.c~a
+++ a/kernel/sys.c
@@ -1713,19 +1713,32 @@ static int validate_prctl_map_locked(str
 			__rc = -EINVAL;					\
 		__rc;							\
 	})
-	error |= __prctl_check_addr_space(start_code);
-	error |= __prctl_check_addr_space(end_code);
-	error |= __prctl_check_addr_space(start_data);
-	error |= __prctl_check_addr_space(end_data);
-	error |= __prctl_check_addr_space(start_stack);
-	error |= __prctl_check_addr_space(start_brk);
-	error |= __prctl_check_addr_space(brk);
-	error |= __prctl_check_addr_space(arg_start);
-	error |= __prctl_check_addr_space(arg_end);
-	error |= __prctl_check_addr_space(env_start);
-	error |= __prctl_check_addr_space(env_end);
-	if (error)
-		goto out;
+	{
+		static const unsigned short offsets[] = {
+			offsetof(struct prctl_mm_map, start_code),
+			offsetof(struct prctl_mm_map, start_code),
+			offsetof(struct prctl_mm_map, end_code),
+			offsetof(struct prctl_mm_map, start_data),
+			offsetof(struct prctl_mm_map, end_data),
+			offsetof(struct prctl_mm_map, start_stack),
+			offsetof(struct prctl_mm_map, start_brk),
+			offsetof(struct prctl_mm_map, brk),
+			offsetof(struct prctl_mm_map, arg_start),
+			offsetof(struct prctl_mm_map, arg_end),
+			offsetof(struct prctl_mm_map, env_start),
+			offsetof(struct prctl_mm_map, env_end),
+		};
+		int i;
+
+		for (i = 0; i < ARRAY_SIZE(offsets); i++) {
+			u64 val = ((u64 *)prctl_map)[offsets[i]];
+
+			if (val < mmap_min_addr || val >= mmap_max_addr) {
+				error = -EINVAL;
+				goto out;
+			}
+		}
+	}

and it saved 400 bytes of text.

But it's a bit hacky.  Can anyone think of anything smarter?