login
Search
This problem was originally tracked down by Brad Spengler. When calling wireless ioctls, if a driver does not correctly validate/shrink iwp->length, the resulting copy_to_user can leak up to 64K of kernel heap contents. It seems that this is triggerable[1] in 2.6.32 at least on ath5k, but I was not able to track down how. The twisty maze of ioctl handlers stumped me. :) Other drivers I checked did not appear to have any problems, but the potential remains. I'm not sure if this patch is the right approach; it was fixed differently[2] in grsecurity. [1] http://forums.grsecurity.net/viewtopic.php?f=3&t=2290&start=0 [2] http://grsecurity.net/~spender/wireless-infoleak-fix2.patch Reported-by: Brad Spengler <spender@grsecurity.net> Signed-off-by: Kees Cook <kees.cook@canonical.com> --- include/net/iw_handler.h | 1 - net/wireless/wext-core.c | 26 ++------------------------ 2 files changed, 2 insertions(+), 25 deletions(-) diff --git a/include/net/iw_handler.h b/include/net/iw_handler.h index 3afdb21..6c81f29 100644 --- a/include/net/iw_handler.h +++ b/include/net/iw_handler.h @@ -277,7 +277,6 @@ #define IW_DESCR_FLAG_EVENT 0x0002 /* Generate an event on SET */ #define IW_DESCR_FLAG_RESTRICT 0x0004 /* GET : request is ROOT only */ /* SET : Omit payload from generated iwevent */ -#define IW_DESCR_FLAG_NOMAX 0x0008 /* GET : no limit on request size */ /* Driver level flags */ #define IW_DESCR_FLAG_WAIT 0x0100 /* Wait for driver event */ diff --git a/net/wireless/wext-core.c b/net/wireless/wext-core.c index 0ef17bc..55b1fd9 100644 --- a/net/wireless/wext-core.c +++ b/net/wireless/wext-core.c @@ -82,7 +82,6 @@ static const struct iw_ioctl_description standard_ioctl[] = { .header_type = IW_HEADER_TYPE_POINT, .token_size = sizeof(struct iw_priv_args), .max_tokens = 16, - .flags = IW_DESCR_FLAG_NOMAX, }, [IW_IOCTL_IDX(SIOCSIWSTATS)] = { .header_type = IW_HEADER_TYPE_NULL, @@ -134,7 +133,6 @@ static const struct iw_ioctl_description ...
Did you tried your patch for real ? With large scan request ?
I ask because at first glance, it looks incorrect, asI believe
I believe this patch would make the situation worse.
Would you mind validating the following patch ? I've just
verified that it compiles and I believe it does what you are asking in
a much more predictable way.
Regards,
Jean
Signed-off-by: Jean Tourrilhes <jt@hpl.hp.com>
diff -u -p wext.j2.c wext.c
--- wext.j2.c 2010-08-27 14:17:26.000000000 -0700
+++ wext.c 2010-08-27 14:19:33.000000000 -0700
@@ -800,9 +800,12 @@ static int ioctl_standard_iw_point(struc
goto out;
}
- if (copy_to_user(iwp->pointer, extra,
- iwp->length *
- descr->token_size)) {
+ /* Verify how much we should return. Some driver
+ * may abuse iwp->length... */
+ if((iwp->length * descr->token_size) < extra_size)
+ extra_size = iwp->length * descr->token_size;
+
+ if (copy_to_user(iwp->pointer, extra, extra_size)) {
err = -EFAULT;
goto out;
}
--
Hi Jean, I did not, no. Since I couldn't reproduce the original problem (I lacked the hardware to test ath5k, and all the other drivers correctly managed The comment should probably be clarified -- it's the caller's iwp->length that may be causing problems (when combined with a driver that forgets to adjusted iwp->length downward). Regardless, the above patch would appear to limit the copy_to_user to only the kzalloced region. Thanks! -Kees -- Kees Cook Ubuntu Security Team --
Ha ! I see. It would be for regular iwpoint queries, not for extended NOMAX queries (scan is a extended NOMAX query). Note that I don't like the idea of reducing the mallocated size, especially with regular queries, as I know that some driver may expect a fixed size in extra and may memcpy to it without double Regards, Jean --
Jean, can you submit in a new thread and right before the SOB add in the commit log Cc: stable@kernel.org [2.6.32+] Luis --
The current patch was made for 2.6.27 and was only compiled. Someone would need to verify it works for 2.6.32. I could Regards, Jean --
Got it, ah so it would be Cc: stable@kernel.org [2.6.27+]. To get this trickled in we first need it for wireless-testing.git, and provide links / patches to the backport of the patch for each kernel. Once it gets merged into Linus' tree the stable team can apply the respective backported patches. Luis --
Based on the code _before_ this hunk, I believe this patch to be wrong
(the goto out matches):
/* If we have something to return to the user */
if (!err && IW_IS_GET(cmd)) {
/* Check if there is enough buffer up there */
if (user_length < iwp->length) {
err = -E2BIG;
goto out;
}
Thus, apparently drivers were intended to be allowed to return more
information than userspace had allocated space for (which also matches
the initial extra_size calculation in this function), so your comment is
wrong, and your check is also wrong because you actually put the burden
on the driver, contrary to the apparent intention of this code.
I believe the below patch is a much better fix as it allows the -E2BIG
code path to be invoked which is more informative to users than
truncated information (which, in your code, may even be truncated in the
middle of a token!!)
johannes
---
net/wireless/wext-core.c | 23 ++++++++++-------------
1 file changed, 10 insertions(+), 13 deletions(-)
--- wireless-testing.orig/net/wireless/wext-core.c 2010-08-30 10:35:33.000000000 +0200
+++ wireless-testing/net/wireless/wext-core.c 2010-08-30 10:46:45.000000000 +0200
@@ -738,26 +738,23 @@ static int ioctl_standard_iw_point(struc
/* Save user space buffer size for checking */
user_length = iwp->length;
- /* Don't check if user_length > max to allow forward
- * compatibility. The test user_length < min is
- * implied by the test at the end.
- */
-
/* Support for very large requests */
- if ((descr->flags & IW_DESCR_FLAG_NOMAX) &&
- (user_length > descr->max_tokens)) {
+ if (descr->flags & IW_DESCR_FLAG_NOMAX) {
/* Allow userspace to GET more than max so
* we can support any size GET requests.
- * There is still a limit : -ENOMEM.
- */
- extra_size = user_length * descr->token_size;
-
- /* Note : user_length is originally a __u16,
+ * ...err, forgot quilt refresh, here's the right patch:
---
net/wireless/wext-core.c | 23 ++++++++++-------------
1 file changed, 10 insertions(+), 13 deletions(-)
--- wireless-testing.orig/net/wireless/wext-core.c 2010-08-30 10:35:33.000000000 +0200
+++ wireless-testing/net/wireless/wext-core.c 2010-08-30 10:57:38.000000000 +0200
@@ -738,26 +738,23 @@ static int ioctl_standard_iw_point(struc
/* Save user space buffer size for checking */
user_length = iwp->length;
- /* Don't check if user_length > max to allow forward
- * compatibility. The test user_length < min is
- * implied by the test at the end.
- */
-
/* Support for very large requests */
- if ((descr->flags & IW_DESCR_FLAG_NOMAX) &&
- (user_length > descr->max_tokens)) {
+ if (descr->flags & IW_DESCR_FLAG_NOMAX) {
/* Allow userspace to GET more than max so
* we can support any size GET requests.
- * There is still a limit : -ENOMEM.
- */
- extra_size = user_length * descr->token_size;
-
- /* Note : user_length is originally a __u16,
+ * There is still a limit: 64k records of
+ * token_size (since a u16 is used).
+ *
+ * Note: user_length is originally a u16,
* and token_size is controlled by us,
* so extra_size won't get negative and
* won't overflow...
*/
- }
+ if (user_length > descr->max_tokens)
+ extra_size = user_length * descr->token_size;
+ } else
+ user_length = min_t(int, user_length,
+ descr->max_tokens);
}
/* kzalloc() ensures NULL-termination for essid_compat. */
--
Ok I finally fully understood the issue. This will fix the problem, but the comment is completely bogus, which I My patch also didn't fix the problem, I didn't understand the problem correctly and was continuously wondering how drivers would ever fill the buffer with more than max_tokens (which would be a more serious bug, since they'd overwrite a slab object after "extra"). What really fixes the problem is the patch below though. Had to realise that the path where the driver didn't do ANYTHING AT ALL was the problem.... johannes Subject: wireless extensions: fix kernel heap content leak From: Johannes Berg <johannes.berg@intel.com> Wireless extensions have an unfortunate, undocumented requirement which requires drivers to always fill iwp->length when returning a successful status. When a driver doesn't do this, it leads to a kernel heap content leak when userspace offers a larger buffer than would have been necessary. Arguably, this is a driver bug, as it should, if it returns 0, fill iwp->length, even if it separately indicated that the buffer contents was not valid. However, we can also remove this requirement and avoid this class of driver bugs by setting the iwp length to max_tokens, which then reflects how big the buffer is that the driver may fill, regardless of how big the userspace buffer is. To illustrate the point, this patch also fixes a corresponding cfg80211 bug (since this requirement isn't documented nor was ever pointed out by anyone during code review, I don't trust all drivers nor all cfg80211 handlers to implement it correctly). Signed-off-by: Johannes Berg <johannes.berg@intel.com> --- net/wireless/wext-compat.c | 3 +++ net/wireless/wext-core.c | 16 ++++++++++++++++ 2 files changed, 19 insertions(+) --- wireless-testing.orig/net/wireless/wext-core.c 2010-08-30 11:29:26.000000000 +0200 +++ wireless-testing/net/wireless/wext-core.c 2010-08-30 11:57:41.000000000 +0200 @@ -782,6 +782,22 @@ static int ...
From: Johannes Berg <johannes.berg@intel.com> Wireless extensions have an unfortunate, undocumented requirement which requires drivers to always fill iwp->length when returning a successful status. When a driver doesn't do this, it leads to a kernel heap content leak when userspace offers a larger buffer than would have been necessary. Arguably, this is a driver bug, as it should, if it returns 0, fill iwp->length, even if it separately indicated that the buffer contents was not valid. However, we can also at least avoid the memory content leak if the driver doesn't do this by setting the iwp length to max_tokens, which then reflects how big the buffer is that the driver may fill, regardless of how big the userspace buffer is. To illustrate the point, this patch also fixes a corresponding cfg80211 bug (since this requirement isn't documented nor was ever pointed out by anyone during code review, I don't trust all drivers nor all cfg80211 handlers to implement it correctly). Cc: stable@kernel.org [all the way back] Signed-off-by: Johannes Berg <johannes.berg@intel.com> --- Tested all four cases now (both hunks of the patch present/not present) and reproduced the problem in the unpatched case. net/wireless/wext-compat.c | 3 +++ net/wireless/wext-core.c | 16 ++++++++++++++++ 2 files changed, 19 insertions(+) --- wireless-testing.orig/net/wireless/wext-core.c 2010-08-30 12:04:57.000000000 +0200 +++ wireless-testing/net/wireless/wext-core.c 2010-08-30 12:10:35.000000000 +0200 @@ -782,6 +782,22 @@ static int ioctl_standard_iw_point(struc } } + if (IW_IS_GET(cmd) && !(descr->flags & IW_DESCR_FLAG_NOMAX)) { + /* + * If this is a GET, but not NOMAX, it means that the extra + * data is not bounded by userspace, but by max_tokens. Thus + * set the length to max_tokens. This matches the extra data + * allocation. + * The driver should fill it with the number of tokens it + * provided, and it may check iwp->length rather than having + ...
Hi Johannes, Thanks for all your work on this! Were you able to trigger the leak through cfg80211? If so, then this will need a CVE assigned and sent to stable too, I think. -Kees -- Kees Cook Ubuntu Security Team --
Yes, I was, very easily, by doing an SIOCGIWESSID while unassociated, with a large iwq->length set by userspace. I did CC stable on my patch, but we can amend the commit by a CVE if so desired. johannes --
Correct, Kees pointed out that my comment was bogus and the
e-mail I sent after the patch corrected myself on that point :
Ha ! I see. It would be for regular iwpoint queries, not for
extended NOMAX queries (scan is a extended NOMAX query).
Yes, I had arrived at the same conclusion (not that my patch
I actually like your patch better than mine, it's closer to
Thanks a lot for the second pair of eyes.
Jean
--
And thank you for looking over my patch. I guess the real bug is still in cfg80211 there. Initially I considered setting iwp->length to zero, rather than max_tokens. What do you think about doing that? The reason I decided to use max_tokens was that somebody might look at the length value to know how much to copy (which is OK only in NOMAX queries); but that would be a more severe driver bug ... can't really make up my mind. We just copy back zeroes anyway. johannes --
Is this fixed differently upstream in the kernel with commit id 42da2f948d949efd0111309f5827bf0298bcc9a4? thanks, greg k-h --
Yes, that's the fix for this. johannes --
Wonderful, thanks for letting me know. greg k-h --
