inet: stop leaking jiffies on the wire
Subsystems (6)
PATCH 1/1 inet: stop leaking jiffies on the wire
Historically linux tried to stick to RFC 791, 1122, 2003 for IPv4 ID field generation.
RFC 6864 made clear that no matter how hard we try, we can not ensure unicity of IP ID within maximum lifetime for all datagrams with a given source address/destination address/protocol tuple.
Linux uses a per socket inet generator (inet_id), initialized at connection startup with a XOR of 'jiffies' and other fields that appear clear on the wire.
Thiemo Nagel pointed that this strategy is a privacy concern as this provides 16 bits of entropy to fingerprint devices.
Let's switch to a random starting point, this is just as good as far as RFC 6864 is concerned and does not leak anything critical.
--- drivers/crypto/chelsio/chtls/chtls_cm.c | 2 +- net/dccp/ipv4.c | 2 +- net/ipv4/datagram.c | 2 +- net/ipv4/tcp_ipv4.c | 4 ++-- net/sctp/socket.c | 2 +- 5 files changed, 6 insertions(+), 6 deletions(-)
drivers/crypto/chelsio/chtls/chtls_cm.c +1 -1
@@ -1297,7 +1297,7 @@ static void make_established(struct sock *sk, u32 snd_isn, unsigned int opt) tp->write_seq = snd_isn; tp->snd_nxt = snd_isn; tp->snd_una = snd_isn; -inet_sk (sk)->inet_id = tp->write_seq ^ jiffies; +inet_sk (sk)->inet_id =prandom_u32 ();assign_rxopt (sk, opt);if (tp->rcv_wnd > (RCV_BUFSIZ_M <<10 ))
net/dccp/ipv4.c +1 -1
@@ -117,7 +117,7 @@ int dccp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len) inet->inet_daddr, inet->inet_sport, inet->inet_dport); - inet->inet_id = dp->dccps_iss ^ jiffies; + inet->inet_id =prandom_u32 (); err =dccp_connect (sk); rt =NULL ;
net/ipv4/datagram.c +1 -1
@@ -73,7 +73,7 @@ int __ip4_datagram_connect(struct sock *sk, struct sockaddr *uaddr, int addr_lenreuseport_has_conns (sk, true); sk->sk_state =TCP_ESTABLISHED ;sk_set_txhash (sk); - inet->inet_id = jiffies; + inet->inet_id =prandom_u32 ();sk_dst_set (sk, &rt->dst); err =0 ;
net/ipv4/tcp_ipv4.c +2 -2
@@ -303,7 +303,7 @@ int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len) inet->inet_daddr); } - inet->inet_id = tp->write_seq ^ jiffies; + inet->inet_id =prandom_u32 ();if (tcp_fastopen_defer_connect (sk, &err))return err;
@@ -1450,7 +1450,7 @@ struct sock *tcp_v4_syn_recv_sock(const struct sock *sk, struct sk_buff *skb,inet_csk (newsk)->icsk_ext_hdr_len =0 ;if (inet_opt)inet_csk (newsk)->icsk_ext_hdr_len = inet_opt->opt.optlen; - newinet->inet_id = newtp->write_seq ^ jiffies; + newinet->inet_id =prandom_u32 ();if (!dst) { dst =inet_csk_route_child_sock (sk, newsk, req);
net/sctp/socket.c +1 -1
@@ -9306,7 +9306,7 @@ void sctp_copy_sock(struct sock *newsk, struct sock *sk, newinet->inet_rcv_saddr = inet->inet_rcv_saddr; newinet->inet_dport =htons (asoc->peer.port); newinet->pmtudisc = inet->pmtudisc; - newinet->inet_id = asoc->next_tsn ^ jiffies; + newinet->inet_id =prandom_u32 (); newinet->uc_ttl = inet->uc_ttl; newinet->mc_loop =1 ;
Signature · 1 line
2.24.0.rc1.363.gb1bccd3e3d-goog
David Miller
maintainerFrom: Eric Dumazet <[email protected]>
Date: Fri, 1 Nov 2019 10:32:19 -0700
Quoted 16 lines
Historically linux tried to stick to RFC 791, 1122, 2003 for IPv4 ID field generation.
RFC 6864 made clear that no matter how hard we try, we can not ensure unicity of IP ID within maximum lifetime for all datagrams with a given source address/destination address/protocol tuple.
Linux uses a per socket inet generator (
inet_id), initialized at connection startup with a XOR of 'jiffies' and other fields that appear clear on the wire.Thiemo Nagel pointed that this strategy is a privacy concern as this provides 16 bits of entropy to fingerprint devices.
Quoted 7 lines
Let's switch to a random starting point, this is just as good as far as RFC 6864 is concerned and does not leak anything critical.
Fixes 1da177e4c3f4 ("Linux-2.6.12-rc2")Signed-off-by Eric Dumazet <[email protected]>Reported-by Thiemo Nagel <[email protected]>
Applied and queued up for -stable, thanks.
Thanks a lot, Eric!
Grepping through the source, it seems to me there are two more occurrences of jiffies in inet_id:
https://github.com/torvalds/linux/blob/v5.3/net/dccp/ipv4.c#L120
https://github.com/torvalds/linux/blob/v5.3/net/dccp/ipv4.c#L419
Kind regards,
Thiemo
Quoted 9 lines · Thiemo Nagel
Thanks a lot, Eric!
Grepping through the source, it seems to me there are two more occurrences of jiffies in
inet_id:https://github.com/torvalds/linux/blob/v5.3/net/dccp/ipv4.c#L120
https://github.com/torvalds/linux/blob/v5.3/net/dccp/ipv4.c#L419
Indeed.
The one in dccp_v4_connect() has been handled in my patch. I missed it in dccp_v4_request_recv_sock()
Thanks.
Quoted 2 lines
Kind regards,
Thiemo
Hey Eric,
I've been thinking about this some more. The prandom_u32() description
says: "This algorithm is NOT considered safe for cryptographic use." -- Afaiu this implies that an attacker could deduce internal state by looking at a sequence of random numbers. Consequently, I believe that we shouldn't use prandom_* for data that gets sent over the wire. Instead get_random_* should be used which is described as cryptographically secure.
Kind regards,
Thiemo
From /drivers/char/random.c:
[About get_random_*:]
* Besides the obvious cryptographic uses, these numbers are also good
* for seeding TCP sequence numbers, and other places where it is
* desirable to have numbers which are not only random, but hard to
* predict by an attacker.
[...]
* It *is* safe to expose get_random_int() output to attackers (e.g. as
* network cookies); given outputs 1..n, it's not feasible to predict
* outputs 0 or n+1.
[...]
* prandom_u32()
* -------------
*
* For even weaker applications, see the pseudorandom generator
* prandom_u32(), prandom_max(), and prandom_bytes(). If the random
* numbers aren't security-critical at all, these are *far* cheaper.
* Useful for self-tests, random error simulation, randomized backoffs,
* and any other application where you trust that nobody is trying to
* maliciously mess with you by guessing the "random" numbers.
Quoted 21 lines · Eric Dumazet
Quoted 9 lines · Thiemo Nagel
Thanks a lot, Eric!
Grepping through the source, it seems to me there are two more occurrences of jiffies in
inet_id:https://github.com/torvalds/linux/blob/v5.3/net/dccp/ipv4.c#L120
https://github.com/torvalds/linux/blob/v5.3/net/dccp/ipv4.c#L419Indeed.
The one in
dccp_v4_connect()has been handled in my patch. I missed it indccp_v4_request_recv_sock()Thanks.
Quoted 2 lines
Kind regards,
Thiemo
Signature · 12 lines
Thiemo Nagel Software Engineer Google Germany GmbH, Erika-Mann-Straße 33, 80686 München Geschäftsführer: Paul Manicle, Halimah DeLaine Prado Registergericht und -nummer: Hamburg, HRB 86891 Sitz der Gesellschaft: Hamburg
Quoted 7 lines · Thiemo Nagel
Hey Eric,
I've been thinking about this some more. The
prandom_u32()description
says: "This algorithm is NOT considered safe for cryptographic use." -- Afaiu this implies that an attacker could deduce internal state by looking at a sequence of random numbers.
Keep in mind that we need some entropy, and some device can not provide any.
prandom_u32() gets some reasonable amount of entropy. And it seems better than ' jiffies' .
If you want strong (hardware) generators, I guess next Android/Chrome devices will cost a bit more.
Consequently, I believe that
Quoted 4 lines
we shouldn't use prandom_* for data that gets sent over the wire. Instead get_random_* should be used which is described as cryptographically secure.
If IP ID had to be cryptographically secure, you can be sure we would have addressed the problem 20 years ago.
Please discuss this matter with random maintainers, not networking ones ;)
Quoted 66 lines
Kind regards,
ThiemoFrom /drivers/char/random.c:
[About get_random_*:]
* Besides the obvious cryptographic uses, these numbers are also good
* for seeding TCP sequence numbers, and other places where it is
* desirable to have numbers which are not only random, but hard to
* predict by an attacker.
[...]
* It *is* safe to exposeget_random_int()output to attackers (e.g. as
* network cookies); given outputs 1..n, it's not feasible to predict
* outputs 0 or n+1.
[...]*
prandom_u32()
* -------------
*
* For even weaker applications, see the pseudorandom generator
*prandom_u32(),prandom_max(), andprandom_bytes(). If the random
* numbers aren't security-critical at all, these are *far* cheaper.
* Useful for self-tests, random error simulation, randomized backoffs,
* and any other application where you trust that nobody is trying to
* maliciously mess with you by guessing the "random" numbers.Quoted 12 lines · Eric Dumazet
Quoted 9 lines · Thiemo Nagel
Thanks a lot, Eric!
Grepping through the source, it seems to me there are two more occurrences of jiffies in
inet_id:https://github.com/torvalds/linux/blob/v5.3/net/dccp/ipv4.c#L120
https://github.com/torvalds/linux/blob/v5.3/net/dccp/ipv4.c#L419Quoted 9 lines
Indeed.
The one in
dccp_v4_connect()has been handled in my patch. I missed it indccp_v4_request_recv_sock()Thanks.
Quoted 2 lines
Kind regards,
Thiemo--
Thiemo Nagel
Software Engineer
Google Germany GmbH, Erika-Mann-Straße 33, 80686 München
Geschäftsführer: Paul Manicle, Halimah DeLaine Prado
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg
Hey Eric,
I think we should distinguish two different problems here:
a) There's insufficient entropy to seed the RNG. As a consequence, observing a (partial) sequence of numbers may allow an attacker to determine the internal state, independently of whether the RNG is cryptographically secure or not.
b) The RNG is not cryptographically secure. In that case, observing a (partial) sequence of numbers may allow an attacker to determine the internal state, independent of the amount of entropy that was used to seed it.
Problem a) is hard -- as you mention, it may require hardware support to solve it fully. However the problem that I'm suggesting to address is b), and that likely can be solved by swapping out prandom_u32() for get_random_u32().
Quoted 2 lines
If IP ID had to be cryptographically secure, you can be sure we would have addressed the problem 20 years ago.
I don't think this is a valid point at all. There are countless examples of things that weren't known 20 years ago but that are better understood today. One relevant example is RFC7258 which only came out in 2014.
Kind regards,
Thiemo
Quoted 5 lines · Thiemo Nagel
Problem a) is hard -- as you mention, it may require hardware support to solve it fully. However the problem that I'm suggesting to address is b), and that likely can be solved by swapping out
prandom_u32()forget_random_u32().
Make sure a server can still generates 10 millions packets per second using this thing :)
I believe get_random_u32() is at least 5 times more expensive than prandom_u32(),
and that is on x86 where arch_has_random() is true and rdrand instruction can be used.
On other arches this would probably kill performance, and eat entropy way too much.
Really this is a discussion that should be taken to random subsystem maintainers.
We can not simply use get_random_u32() here.