Skip to contentSkip to conversation

x86, x32: Correct invalid use of user timespec in the kernel

DISCUSSION Dave Jones 4 participants 7 messages Jan 31, 2014 – Feb 1, 2014 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Dave Jones

author
#1 lore raw
Quoted 5 lines · Linux Kernel

Commit: 2def2ef2ae5f3990aabdbe8a755911902707d268

...

Review

Quoted 9 lines
(quoted diff) +2 -2
-	if (get_compat_timespec(&ktspec, timeout))
+	if (compat_get_timespec(&ktspec, timeout))
 		return -EFAULT;
 
 	datagrams = __sys_recvmmsg(fd, (struct mmsghdr __user *)mmsg, vlen,
 				   flags | MSG_CMSG_COMPAT, &ktspec);
-	if (datagrams > 0 && put_compat_timespec(&ktspec, timeout))
+	if (datagrams > 0 && compat_put_timespec(&ktspec, timeout))
 		datagrams = -EFAULT;

Can we rename one of each of those functions ?
It's not really surprising they got mixed up given they look so alike.

It looks like an accident just waiting to happen again.

Dave

#2 lore raw
Quoted 20 lines · Dave Jones
Quoted 5 lines · Linux Kernel

Commit: 2def2ef2ae5f3990aabdbe8a755911902707d268

...

Review

Quoted 9 lines
(quoted diff) +2 -2
-	if (get_compat_timespec(&ktspec, timeout))
+	if (compat_get_timespec(&ktspec, timeout))
 		return -EFAULT;
 
 	datagrams = __sys_recvmmsg(fd, (struct mmsghdr __user *)mmsg, vlen,
 				   flags | MSG_CMSG_COMPAT, &ktspec);
-	if (datagrams > 0 && put_compat_timespec(&ktspec, timeout))
+	if (datagrams > 0 && compat_put_timespec(&ktspec, timeout))
 		datagrams = -EFAULT;

Can we rename one of each of those functions ?
It's not really surprising they got mixed up given they look so alike.

Quoted 2 lines

It looks like an accident just waiting to happen again.

Very much so, I made the same comment.

My feeling is that {get,put}_compat_timespec() should at the very least have leading underscores to flag it as a low-level function, but better suggestions would be appreciated.

-hpa

#3 lore raw
Quoted 4 lines · H. Peter Anvin

My feeling is that {get,put}_compat_timespec() should at the very least have leading underscores to flag it as a low-level function, but better suggestions would be appreciated.

Why not just remove it entirely, and change all users to compat_[get|set]_timespec (same for timeval etc, of course).

After all, compat_*_time*() does fall back cleanly for non-x32 cases. And sure, maybe that particular code is never *needed* for x32 support, but the overhead is generally zero (since in most cases X32 isn't even configured), or very low anyway. So the upside of having two subtly incompatible interfaces is very dubious, no?

Linus

#4 lore raw
Quoted 6 lines · Linus Torvalds
Quoted 4 lines · H. Peter Anvin

My feeling is that {get,put}_compat_timespec() should at the very least have leading underscores to flag it as a low-level function, but better suggestions would be appreciated.

Quoted 9 lines

Why not just remove it entirely, and change all users to compat_[get|set]_timespec (same for timeval etc, of course).

After all, compat_*_time*() does fall back cleanly for non-x32 cases. And sure, maybe that particular code is never *needed* for x32 support, but the overhead is generally zero (since in most cases X32 isn't even configured), or very low anyway. So the upside of having two subtly incompatible interfaces is very dubious, no?

As they both seem to be out of line, I would think so. More than half of the use cases are in kernel/compat.c where we could use a double-underscore inline version if we really care -- it would probably be a net win in terms of performance.

There are only 25 call sites in the kernel of '(get|put)_compat_time(val|spec)' and that includes the ones inside the larger functions.

-hpa

#5 lore raw
In reply to Linus Torvalds
Quoted 6 lines · Linus Torvalds
Quoted 4 lines · H. Peter Anvin

My feeling is that {get,put}_compat_timespec() should at the very least have leading underscores to flag it as a low-level function, but better suggestions would be appreciated.

Quoted 9 lines

Why not just remove it entirely, and change all users to compat_[get|set]_timespec (same for timeval etc, of course).

After all, compat_*_time*() does fall back cleanly for non-x32 cases. And sure, maybe that particular code is never *needed* for x32 support, but the overhead is generally zero (since in most cases X32 isn't even configured), or very low anyway. So the upside of having two subtly incompatible interfaces is very dubious, no?

Hmmm... it ends up being a bit weird even so. Some of the interfaces ought to be revamped at a higher level.

Consider this bit in ipc/compat.c:

long compat_sys_semtimedop(int semid, struct sembuf __user *tsems,
unsigned nsops, const struct compat_timespec __user
*timeout)

{
        struct timespec __user *ts64 = NULL;
        if (timeout) {
                struct timespec ts;
                ts64 = compat_alloc_user_space(sizeof(*ts64));
                if (get_compat_timespec(&ts, timeout))
                        return -EFAULT;
                if (copy_to_user(ts64, &ts, sizeof(ts)))
                        return -EFAULT;
        }
        return sys_semtimedop(semid, tsems, nsops, ts64);
}

This is most definitely broken if COMPAT_USE_64BIT_TIME, even with get_compat_timespec() is replaced by compat_get_timespec(). However, what is *really* going on here is that we want to provide a user space pointer to a kernel-format timespec, so we could have an interface like this:

int compat_convert_timespec_user(struct compat_timespec **ts64p, const struct compat_timespec __user *ts)

{
        struct timespec __user *ts64;
        struct timespec ts;

	/* If the compat timespec is 64 bits, no conversion is needed */
	if (!ts || COMPAT_USE_64BIT_TIME) {
		*ts64p = (struct timespec __user *)ts;
		return 0;
	}

	*ts64p = ts64 = compat_alloc_user_space(sizeof(*ts64));
        if (__get_compat_timespec(&ts, timeout))
		return -EFAULT;
	if (copy_to_user(ts64, &ts, sizeof(ts)))
        	return -EFAULT;

	return 0;
}

Now one can argue we have a potential problem with type safety here, but I'm not sure there is any way to avoid that.

-hpa

#6 lore raw
In reply to Linus Torvalds
Quoted 6 lines · Linus Torvalds
Quoted 4 lines · H. Peter Anvin

My feeling is that {get,put}_compat_timespec() should at the very least have leading underscores to flag it as a low-level function, but better suggestions would be appreciated.

Quoted 9 lines

Why not just remove it entirely, and change all users to compat_[get|set]_timespec (same for timeval etc, of course).

After all, compat_*_time*() does fall back cleanly for non-x32 cases. And sure, maybe that particular code is never *needed* for x32 support, but the overhead is generally zero (since in most cases X32 isn't even configured), or very low anyway. So the upside of having two subtly incompatible interfaces is very dubious, no?

Here is a patch for comments/review/opinions... it has only been compile-tested.

-hpa

drivers/media/v4l2-core/v4l2-compat-ioctl32.c +1 -1
index 8f7a6a4..6191968 100644
@@ -733,7 +733,7 @@ static int put_v4l2_event32(struct v4l2_event *kp, struct v4l2_event32 __user *u
 		copy_to_user(&up->u, &kp->u, sizeof(kp->u)) ||
 		put_user(kp->pending, &up->pending) ||
 		put_user(kp->sequence, &up->sequence) ||
-		put_compat_timespec(&kp->timestamp, &up->timestamp) ||
+		compat_put_timespec(&kp->timestamp, &up->timestamp) ||
 		put_user(kp->id, &up->id) ||
 		copy_to_user(up->reserved, kp->reserved, 8 * sizeof(__u32)))
 			return -EFAULT;
fs/compat.c +3 -3
index 6af20de..62092f4 100644
@@ -92,8 +92,8 @@ asmlinkage long compat_sys_utimensat(unsigned int dfd, const char __user *filena
 	struct timespec tv[2];
 
 	if  (t) {
-		if (get_compat_timespec(&tv[0], &t[0]) ||
-		    get_compat_timespec(&tv[1], &t[1]))
+		if (compat_get_timespec(&tv[0], &t[0]) ||
+		    compat_get_timespec(&tv[1], &t[1]))
 			return -EFAULT;
 
 		if (tv[0].tv_nsec == UTIME_OMIT && tv[1].tv_nsec == UTIME_OMIT)
@@ -512,7 +512,7 @@ compat_sys_io_getevents(aio_context_t ctx_id,
 				nr * sizeof(struct io_event))))
 		goto out;
 	if (timeout) {
-		if (get_compat_timespec(&t, timeout))
+		if (compat_get_timespec(&t, timeout))
 			goto out;
 
 		ut = compat_alloc_user_space(sizeof(*ut));
include/linux/compat.h +10 -13
index 19f6003..db467cf 100644
@@ -141,26 +141,23 @@ struct compat_sigaction {
 };
 
 /*
- * These functions operate strictly on struct compat_time*
- */
-extern int get_compat_timespec(struct timespec *,
-			       const struct compat_timespec __user *);
-extern int put_compat_timespec(const struct timespec *,
-			       struct compat_timespec __user *);
-extern int get_compat_timeval(struct timeval *,
-			      const struct compat_timeval __user *);
-extern int put_compat_timeval(const struct timeval *,
-			      struct compat_timeval __user *);
-/*
  * These functions operate on 32- or 64-bit specs depending on
- * COMPAT_USE_64BIT_TIME, hence the void user pointer arguments and the
- * naming as compat_get/put_ rather than get/put_compat_.
+ * COMPAT_USE_64BIT_TIME, hence the void user pointer arguments.
  */
 extern int compat_get_timespec(struct timespec *, const void __user *);
 extern int compat_put_timespec(const struct timespec *, void __user *);
 extern int compat_get_timeval(struct timeval *, const void __user *);
 extern int compat_put_timeval(const struct timeval *, void __user *);
 
+/*
+ * These functions convert a timeval/timespec if necessary and return
+ * a user space pointer.
+ */
+extern int compat_convert_timespec(const struct timespec __user **,
+				   const void __user *);
+extern int compat_convert_timeval(const struct timeval __user **,
+				  const void __user *);
+
 struct compat_iovec {
 	compat_uptr_t	iov_base;
 	compat_size_t	iov_len;
ipc/compat.c +3 -9
index f486b00..1048522 100644
@@ -752,14 +752,8 @@ long compat_sys_shmctl(int first, int second, void __user *uptr)
 long compat_sys_semtimedop(int semid, struct sembuf __user *tsems,
 		unsigned nsops, const struct compat_timespec __user *timeout)
 {
-	struct timespec __user *ts64 = NULL;
-	if (timeout) {
-		struct timespec ts;
-		ts64 = compat_alloc_user_space(sizeof(*ts64));
-		if (get_compat_timespec(&ts, timeout))
-			return -EFAULT;
-		if (copy_to_user(ts64, &ts, sizeof(ts)))
-			return -EFAULT;
-	}
+	struct timespec __user *ts64;
+	if (compat_convert_timespec(&ts64, timeout))
+		return -EFAULT;
 	return sys_semtimedop(semid, tsems, nsops, ts64);
 }
ipc/compat_mq.c +3 -16
index 63d7c6de..a9cf163 100644
@@ -64,20 +64,6 @@ asmlinkage long compat_sys_mq_open(const char __user *u_name,
 	return sys_mq_open(u_name, oflag, mode, p);
 }
 
-static int compat_prepare_timeout(struct timespec __user **p,
-				  const struct compat_timespec __user *u)
-{
-	struct timespec ts;
-	if (!u) {
-		*p = NULL;
-		return 0;
-	}
-	*p = compat_alloc_user_space(sizeof(ts));
-	if (get_compat_timespec(&ts, u) || copy_to_user(*p, &ts, sizeof(ts)))
-		return -EFAULT;
-	return 0;
-}
-
 asmlinkage long compat_sys_mq_timedsend(mqd_t mqdes,
 			const char __user *u_msg_ptr,
 			size_t msg_len, unsigned int msg_prio,
@@ -85,7 +71,7 @@ asmlinkage long compat_sys_mq_timedsend(mqd_t mqdes,
 {
 	struct timespec __user *u_ts;
 
-	if (compat_prepare_timeout(&u_ts, u_abs_timeout))
+	if (compat_convert_timespec(&u_ts, u_abs_timeout))
 		return -EFAULT;
 
 	return sys_mq_timedsend(mqdes, u_msg_ptr, msg_len,
@@ -98,7 +84,8 @@ asmlinkage ssize_t compat_sys_mq_timedreceive(mqd_t mqdes,
 			const struct compat_timespec __user *u_abs_timeout)
 {
 	struct timespec __user *u_ts;
-	if (compat_prepare_timeout(&u_ts, u_abs_timeout))
+
+	if (compat_convert_timespec(&u_ts, u_abs_timeout))
 		return -EFAULT;
 
 	return sys_mq_timedreceive(mqdes, u_msg_ptr, msg_len,
kernel/compat.c +77 -54

281 lines not on this page · raw message

kernel/futex_compat.c +1 -1
index f9f44fd..55c8c93 100644
@@ -183,7 +183,7 @@ COMPAT_SYSCALL_DEFINE6(futex, u32 __user *, uaddr, int, op, u32, val,
 	if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI ||
 		      cmd == FUTEX_WAIT_BITSET ||
 		      cmd == FUTEX_WAIT_REQUEUE_PI)) {
-		if (get_compat_timespec(&ts, utime))
+		if (compat_get_timespec(&ts, utime))
 			return -EFAULT;
 		if (!timespec_valid(&ts))
 			return -EINVAL;
#7 lore raw
Quoted 3 lines · H. Peter Anvin

Here is a patch for comments/review/opinions... it has only been compile-tested.

Looks sane to me from a superficial quick read-through.

Linus