[PATCH 0/3] xfrm/compat: syzbot-found fixes

WARN_ON() for XFRMA_UNSPEC translation which likely no-one except
syzkaller uses; properly zerofy tail-padding for 64-bit attribute;
don't use __GFP_ZERO as the memory is initialized during translation.

Cc: Steffen Klassert <steffen.klassert@secunet.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Jakub Kicinski <kuba@kernel.org>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Hillf Danton <redacted>
Cc: <redacted>

Thanks,
         Dmitry

Dmitry Safonov (3):
  xfrm/compat: Translate by copying XFRMA_UNSPEC attribute
  xfrm/compat: memset(0) 64-bit padding at right place
  xfrm/compat: Don't allocate memory with __GFP_ZERO

 net/xfrm/xfrm_compat.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)


base-commit: 07e0887302450a62f51dba72df6afb5fabb23d1c
-- 
2.28.0

xfrm_xlate32() translates 64-bit message provided by kernel to be sent
for 32-bit listener (acknowledge or monitor). Translator code doesn't
expect XFRMA_UNSPEC attribute as it doesn't know its payload.
Kernel never attaches such attribute, but a user can.

I've searched if any opensource does it and the answer is no.
Nothing on github and google finds only tfcproject that has such code
commented-out.

What will happen if a user sends a netlink message with XFRMA_UNSPEC
attribute? Ipsec code ignores this attribute. But if there is a
monitor-process or 32-bit user requested ack - kernel will try to
translate such message and will hit WARN_ONCE() in xfrm_xlate64_attr().

Deal with XFRMA_UNSPEC by copying the attribute payload with
xfrm_nla_cpy(). In result, the default switch-case in xfrm_xlate64_attr()
becomes an unused code. Leave those 3 lines in case a new xfrm attribute
will be added.

Reported-by: <redacted>
Signed-off-by: Dmitry Safonov <redacted>
---
 net/xfrm/xfrm_compat.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/net/xfrm/xfrm_compat.c b/net/xfrm/xfrm_compat.c
index e28f0c9ecd6a..17edbf935e35 100644
--- a/net/xfrm/xfrm_compat.c
+++ b/net/xfrm/xfrm_compat.c

@@ -234,6 +234,7 @@ static int xfrm_xlate64_attr(struct sk_buff *dst, const struct nlattr *src)
 	case XFRMA_PAD:
 		/* Ignore */
 		return 0;
+	case XFRMA_UNSPEC:
 	case XFRMA_ALG_AUTH:
 	case XFRMA_ALG_CRYPT:
 	case XFRMA_ALG_COMP:

-- 
2.28.0

32-bit messages translated by xfrm_compat can have attributes attached.
For all, but XFRMA_SA, XFRMA_POLICY the size of payload is the same
in 32-bit UABI and 64-bit UABI. For XFRMA_SA (struct xfrm_usersa_info)
and XFRMA_POLICY (struct xfrm_userpolicy_info) it's only tail-padding
that is present in 64-bit payload, but not in 32-bit.
The proper size for destination nlattr is already calculated by
xfrm_user_rcv_calculate_len64() and allocated with kvmalloc().

xfrm_attr_cpy32() copies 32-bit copy_len into 64-bit attribute
translated payload, zero-filling possible padding for SA/POLICY.
Due to a typo, *pos already has 64-bit payload size, in a result next
memset(0) is called on the memory after the translated attribute, not on
the tail-padding of it.

Reported-by: <redacted>
Signed-off-by: Dmitry Safonov <redacted>
---
 net/xfrm/xfrm_compat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/net/xfrm/xfrm_compat.c b/net/xfrm/xfrm_compat.c
index 17edbf935e35..556e9f33b815 100644
--- a/net/xfrm/xfrm_compat.c
+++ b/net/xfrm/xfrm_compat.c

@@ -388,7 +388,7 @@ static int xfrm_attr_cpy32(void *dst, size_t *pos, const struct nlattr *src,
 
 	memcpy(nla, src, nla_attr_size(copy_len));
 	nla->nla_len = nla_attr_size(payload);
-	*pos += nla_attr_size(payload);
+	*pos += nla_attr_size(copy_len);
 	nlmsg->nlmsg_len += nla->nla_len;
 
 	memset(dst + *pos, 0, payload - copy_len);

-- 
2.28.0

32-bit to 64-bit messages translator zerofies needed paddings in the
translation, the rest is the actual payload.
Don't allocate zero pages as they are not needed.

Signed-off-by: Dmitry Safonov <redacted>
---
 net/xfrm/xfrm_compat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/net/xfrm/xfrm_compat.c b/net/xfrm/xfrm_compat.c
index 556e9f33b815..d8e8a11ca845 100644
--- a/net/xfrm/xfrm_compat.c
+++ b/net/xfrm/xfrm_compat.c

@@ -564,7 +564,7 @@ static struct nlmsghdr *xfrm_user_rcv_msg_compat(const struct nlmsghdr *h32,
 		return NULL;
 
 	len += NLMSG_HDRLEN;
-	h64 = kvmalloc(len, GFP_KERNEL | __GFP_ZERO);
+	h64 = kvmalloc(len, GFP_KERNEL);
 	if (!h64)
 		return ERR_PTR(-ENOMEM);
 

-- 
2.28.0

On Fri, Oct 30, 2020 at 02:25:57AM +0000, Dmitry Safonov wrote:

WARN_ON() for XFRMA_UNSPEC translation which likely no-one except
syzkaller uses; properly zerofy tail-padding for 64-bit attribute;
don't use __GFP_ZERO as the memory is initialized during translation.

Cc: Steffen Klassert <steffen.klassert@secunet.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Jakub Kicinski <kuba@kernel.org>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Hillf Danton <redacted>
Cc: <redacted>

Thanks,
         Dmitry

Dmitry Safonov (3):
  xfrm/compat: Translate by copying XFRMA_UNSPEC attribute
  xfrm/compat: memset(0) 64-bit padding at right place
  xfrm/compat: Don't allocate memory with __GFP_ZERO

Can you please add 'Fixes' tags to all the patches.

Thanks!