From: Dmitry Safonov <0x7f454c46@gmail.com>
Currently TCP-MD5 keys are dumped as an array of
(struct tcp_diag_md5sig). All the keys from a socket go
into the same netlink attribute. The maximum amount of TCP-MD5 keys on
any socket is limited by /proc/sys/net/core/optmem_max, which post
commit 4944566706b2 ("net: increase optmem_max default value") is now by
default 128 KB. With the help of selftest I've figured out that equals
to 963 keys, without user having to increase optmem_max:
> test_set_md5() [963/1024]: Cannot allocate memory
The maximum length of nlattr is limited by typeof(nlattr::nla_len),
which is (U16_MAX - 1). When there are too many keys the array written
overflows the netlink attribute. Here is what one can see on a test,
with no adjustments to optmem_max defaults:
> recv() = 65180
> socket: 10.0.254.1:7013->0.0.0.0:0 (intf 3)
> family: 2 state: 10 timer: 0 retrans: 0
> expires: 0 rqueu: 0 wqueue: 1 uid: 0 inode: 456
> attr type: 8 (5)
> attr type: 15 (8)
> attr type: 21 (12)
> attr type: 22 (6)
> attr type: 2 (252)
> attr type: 18 (64804)
> recv() = 130680
> socket: 10.0.254.1:7013->0.0.0.0:0 (intf 3)
> family: 2 state: 10 timer: 0 retrans: 0
> expires: 0 rqueu: 0 wqueue: 1 uid: 0 inode: 456
> attr type: 8 (5)
> attr type: 15 (8)
> attr type: 21 (12)
> attr type: 22 (6)
> attr type: 2 (252)
> attr type: 18 (64768)
> attr type: 29555 (25966)
> recv() = 130680
> socket: 10.0.254.1:7013->0.0.0.0:0 (intf 3)
> family: 2 state: 10 timer: 0 retrans: 0
> expires: 0 rqueu: 0 wqueue: 1 uid: 0 inode: 456
> attr type: 8 (5)
> attr type: 15 (8)
> attr type: 21 (12)
> attr type: 22 (6)
> attr type: 2 (252)
> attr type: 18 (64768)
> attr type: 29555 (25966)
> attr type: 8265 (8236)
Here attribute type 18 is INET_DIAG_MD5SIG, the following nlattr types
are junk made of tcp_diag_md5sig's content.
Here is the overflow of the nlattr size:
>>> hex(64768)
'0xfd00'
>>> hex(130300)
'0x1fcfc'
Limit the size of (struct tcp_diag_md5sig) array in the netlink reply by
maximum attribute length. Not perfect as NLM_F_DUMP_INTR will be set on
the netlink header flags, but the userspace can differ if it's due to
inconsistency or due to maximum size of the netlink attribute.
In a following patch set, I'm planning to address this and re-introduce
TCP-MD5-diag that actually works.
Signed-off-by: Dmitry Safonov <0x7f454c46@gmail.com>
---
net/ipv4/tcp_diag.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/net/ipv4/tcp_diag.c b/net/ipv4/tcp_diag.c
index 722dbfd54d247b4def1e77b1674c5b207c5a939d..d55a0ac39fa0853806efd4a6b38591255e0f4930 100644
--- a/net/ipv4/tcp_diag.c
+++ b/net/ipv4/tcp_diag.c
@@ -72,6 +72,7 @@ static int tcp_diag_put_md5sig(struct sk_buff *skb,
return 0;
attrlen = skb_availroom(skb) - NLA_HDRLEN;
+ attrlen = min(attrlen, U16_MAX - 1); /* attr->nla_len */
md5sig_count = min(md5sig_count, attrlen / key_size);
attr = nla_reserve(skb, INET_DIAG_MD5SIG, md5sig_count * key_size);
if (!attr)
--
2.42.2
From: Dmitry Safonov via B4 Relay <devnull+0x7f454c46.gmail.com@kernel.org> Date: Wed, 06 Nov 2024 18:10:18 +0000 > From: Dmitry Safonov <0x7f454c46@gmail.com> > > Currently TCP-MD5 keys are dumped as an array of > (struct tcp_diag_md5sig). All the keys from a socket go > into the same netlink attribute. The maximum amount of TCP-MD5 keys on > any socket is limited by /proc/sys/net/core/optmem_max, which post > commit 4944566706b2 ("net: increase optmem_max default value") is now by > default 128 KB. With the help of selftest I've figured out that equals > to 963 keys, without user having to increase optmem_max: > > test_set_md5() [963/1024]: Cannot allocate memory > > The maximum length of nlattr is limited by typeof(nlattr::nla_len), > which is (U16_MAX - 1). When there are too many keys the array written > overflows the netlink attribute. Here is what one can see on a test, > with no adjustments to optmem_max defaults: > > > recv() = 65180 > > socket: 10.0.254.1:7013->0.0.0.0:0 (intf 3) > > family: 2 state: 10 timer: 0 retrans: 0 > > expires: 0 rqueu: 0 wqueue: 1 uid: 0 inode: 456 > > attr type: 8 (5) > > attr type: 15 (8) > > attr type: 21 (12) > > attr type: 22 (6) > > attr type: 2 (252) > > attr type: 18 (64804) > > recv() = 130680 > > socket: 10.0.254.1:7013->0.0.0.0:0 (intf 3) > > family: 2 state: 10 timer: 0 retrans: 0 > > expires: 0 rqueu: 0 wqueue: 1 uid: 0 inode: 456 > > attr type: 8 (5) > > attr type: 15 (8) > > attr type: 21 (12) > > attr type: 22 (6) > > attr type: 2 (252) > > attr type: 18 (64768) > > attr type: 29555 (25966) > > recv() = 130680 > > socket: 10.0.254.1:7013->0.0.0.0:0 (intf 3) > > family: 2 state: 10 timer: 0 retrans: 0 > > expires: 0 rqueu: 0 wqueue: 1 uid: 0 inode: 456 > > attr type: 8 (5) > > attr type: 15 (8) > > attr type: 21 (12) > > attr type: 22 (6) > > attr type: 2 (252) > > attr type: 18 (64768) > > attr type: 29555 (25966) > > attr type: 8265 (8236) > > Here attribute type 18 is INET_DIAG_MD5SIG, the following nlattr types > are junk made of tcp_diag_md5sig's content. > > Here is the overflow of the nlattr size: > >>> hex(64768) > '0xfd00' > >>> hex(130300) > '0x1fcfc' > > Limit the size of (struct tcp_diag_md5sig) array in the netlink reply by > maximum attribute length. Not perfect as NLM_F_DUMP_INTR will be set on > the netlink header flags, but the userspace can differ if it's due to > inconsistency or due to maximum size of the netlink attribute. > > In a following patch set, I'm planning to address this and re-introduce > TCP-MD5-diag that actually works. Given the issue has not been reported so far (I think), we can wait for the series rather than backporting this.
On Thu, 7 Nov 2024 at 00:26, Kuniyuki Iwashima <kuniyu@amazon.com> wrote: > > From: Dmitry Safonov via B4 Relay <devnull+0x7f454c46.gmail.com@kernel.org> > Date: Wed, 06 Nov 2024 18:10:18 +0000 > > From: Dmitry Safonov <0x7f454c46@gmail.com> > > > > Currently TCP-MD5 keys are dumped as an array of > > (struct tcp_diag_md5sig). All the keys from a socket go > > into the same netlink attribute. The maximum amount of TCP-MD5 keys on > > any socket is limited by /proc/sys/net/core/optmem_max, which post > > commit 4944566706b2 ("net: increase optmem_max default value") is now by > > default 128 KB. With the help of selftest I've figured out that equals > > to 963 keys, without user having to increase optmem_max: > > > test_set_md5() [963/1024]: Cannot allocate memory > > > > The maximum length of nlattr is limited by typeof(nlattr::nla_len), > > which is (U16_MAX - 1). When there are too many keys the array written > > overflows the netlink attribute. Here is what one can see on a test, > > with no adjustments to optmem_max defaults: > > > > > recv() = 65180 > > > socket: 10.0.254.1:7013->0.0.0.0:0 (intf 3) > > > family: 2 state: 10 timer: 0 retrans: 0 > > > expires: 0 rqueu: 0 wqueue: 1 uid: 0 inode: 456 > > > attr type: 8 (5) > > > attr type: 15 (8) > > > attr type: 21 (12) > > > attr type: 22 (6) > > > attr type: 2 (252) > > > attr type: 18 (64804) > > > recv() = 130680 > > > socket: 10.0.254.1:7013->0.0.0.0:0 (intf 3) > > > family: 2 state: 10 timer: 0 retrans: 0 > > > expires: 0 rqueu: 0 wqueue: 1 uid: 0 inode: 456 > > > attr type: 8 (5) > > > attr type: 15 (8) > > > attr type: 21 (12) > > > attr type: 22 (6) > > > attr type: 2 (252) > > > attr type: 18 (64768) > > > attr type: 29555 (25966) > > > recv() = 130680 > > > socket: 10.0.254.1:7013->0.0.0.0:0 (intf 3) > > > family: 2 state: 10 timer: 0 retrans: 0 > > > expires: 0 rqueu: 0 wqueue: 1 uid: 0 inode: 456 > > > attr type: 8 (5) > > > attr type: 15 (8) > > > attr type: 21 (12) > > > attr type: 22 (6) > > > attr type: 2 (252) > > > attr type: 18 (64768) > > > attr type: 29555 (25966) > > > attr type: 8265 (8236) > > > > Here attribute type 18 is INET_DIAG_MD5SIG, the following nlattr types > > are junk made of tcp_diag_md5sig's content. > > > > Here is the overflow of the nlattr size: > > >>> hex(64768) > > '0xfd00' > > >>> hex(130300) > > '0x1fcfc' > > > > Limit the size of (struct tcp_diag_md5sig) array in the netlink reply by > > maximum attribute length. Not perfect as NLM_F_DUMP_INTR will be set on > > the netlink header flags, but the userspace can differ if it's due to > > inconsistency or due to maximum size of the netlink attribute. > > > > In a following patch set, I'm planning to address this and re-introduce > > TCP-MD5-diag that actually works. > > Given the issue has not been reported so far (I think), we can wait for > the series rather than backporting this. Yeah, my concern is that ss or or other tools may interrupt the md5keys from overflow as other netlink attributes and either show non-meaningful things or even hide some socket information (as if an attribute is met the second time, from what I read in ss code, it will put the last met attribute of the same type over previous pointers and print only that). Regarding reports, one has to have U16_MAX / sizeof(struct tcp_diag_md5sig) = 655 keys on a socket. Probably, not many people use BGP with that many peers. I'm fine moving that to later net-next patches; I've sent it only because the above seemed concerning to me. Thanks, Dmitry
© 2016 - 2024 Red Hat, Inc.