2003-08-27 21:30:10

by Ricky Beam

[permalink] [raw]
Subject: lost socket

It would appear the kernel is not properly reporting all the active sockets.
I noticed this in 2.5.70 while checking netflow collection... udp port 9800
is bound and actively receiving traffic, however, nothing reports it's
existance. 2.6.0-test4 still has this bug.

This smells like a simple "off by one" bug, but I've been too busy to go
look at the code.

--Ricky



2003-08-28 00:01:17

by Ricky Beam

[permalink] [raw]
Subject: /proc/net/* read drops data (was: lost socket)

On Wed, 27 Aug 2003, Ricky Beam wrote:
>This smells like a simple "off by one" bug, but I've been too busy to go
>look at the code.

Ah hah! it's a block size problem... netstat reads 1024 at a time.

Using dd...

[root:pts/5{9}]gir:~/[7:55pm]:dd if=/proc/net/udp bs=1024 | wc
2+1 records in
2+1 records out
18 216 2304
[root:pts/5{9}]gir:~/[7:56pm]:dd if=/proc/net/udp bs=2048 | wc
1+1 records in
1+1 records out
19 228 2432
[root:pts/5{9}]gir:~/[7:56pm]:dd if=/proc/net/udp bs=4096 | wc
0+1 records in
0+1 records out
20 240 2560
[root:pts/5{9}]gir:~/[7:56pm]:dd if=/proc/net/udp bs=8192 | wc
0+1 records in
0+1 records out
20 240 2560
[root:pts/5{9}]gir:~/[7:56pm]:dd if=/proc/net/udp bs=32768 | wc
0+1 records in
0+1 records out
20 240 2560
[root:pts/5{9}]gir:~/[7:56pm]:dd if=/proc/net/udp bs=1 | wc
2432+0 records in
2432+0 records out
19 228 2432

--Ricky


2003-08-28 00:34:06

by YOSHIFUJI Hideaki

[permalink] [raw]
Subject: Re: /proc/net/* read drops data

In article <[email protected]> (at Wed, 27 Aug 2003 19:58:17 -0400 (EDT)), Ricky Beam <[email protected]> says:

> On Wed, 27 Aug 2003, Ricky Beam wrote:
> >This smells like a simple "off by one" bug, but I've been too busy to go
> >look at the code.
>
> Ah hah! it's a block size problem... netstat reads 1024 at a time.
>
> Using dd...
>
> [root:pts/5{9}]gir:~/[7:55pm]:dd if=/proc/net/udp bs=1024 | wc
> 2+1 records in
> 2+1 records out
> 18 216 2304

Good idea. I'll chase this bug.

--yoshfuji

2003-09-03 07:07:25

by YOSHIFUJI Hideaki

[permalink] [raw]
Subject: Re: /proc/net/* read drops data

Hello.

In article <[email protected]> (at Wed, 27 Aug 2003 19:58:17 -0400 (EDT)), Ricky Beam <[email protected]> says:

> On Wed, 27 Aug 2003, Ricky Beam wrote:
> >This smells like a simple "off by one" bug, but I've been too busy to go
> >look at the code.
>
> Ah hah! it's a block size problem... netstat reads 1024 at a time.
>
> Using dd...
>
> [root:pts/5{9}]gir:~/[7:55pm]:dd if=/proc/net/udp bs=1024 | wc
> 2+1 records in
> 2+1 records out
> 18 216 2304
> [root:pts/5{9}]gir:~/[7:56pm]:dd if=/proc/net/udp bs=2048 | wc
> 1+1 records in
> 1+1 records out
> 19 228 2432
:

Please try this patch.

D: Fixing a bug that reading /proc/net/{udp,udp6} may drop some data

Index: linux-2.6/net/ipv4/udp.c
===================================================================
RCS file: /home/cvs/linux-2.5/net/ipv4/udp.c,v
retrieving revision 1.38
diff -u -r1.38 udp.c
--- linux-2.6/net/ipv4/udp.c 14 Aug 2003 06:00:04 -0000 1.38
+++ linux-2.6/net/ipv4/udp.c 3 Sep 2003 05:30:52 -0000
@@ -1349,64 +1349,65 @@
/* ------------------------------------------------------------------------ */
#ifdef CONFIG_PROC_FS

-static __inline__ struct sock *udp_get_bucket(struct seq_file *seq, loff_t *pos)
+static struct sock *udp_get_first(struct seq_file *seq)
{
- int i;
struct sock *sk;
- struct hlist_node *node;
- loff_t l = *pos;
struct udp_iter_state *state = seq->private;

- for (; state->bucket < UDP_HTABLE_SIZE; ++state->bucket) {
- i = 0;
+ for (state->bucket = 0; state->bucket < UDP_HTABLE_SIZE; ++state->bucket) {
+ struct hlist_node *node;
sk_for_each(sk, node, &udp_hash[state->bucket]) {
- if (sk->sk_family != state->family) {
- ++i;
- continue;
- }
- if (l--) {
- ++i;
- continue;
- }
- *pos = i;
- goto out;
+ if (sk->sk_family == state->family)
+ goto found;
}
}
sk = NULL;
-out:
+found:
return sk;
}

+static struct sock *udp_get_next(struct seq_file *seq, struct sock *sk)
+{
+ struct udp_iter_state *state = seq->private;
+
+ do {
+ sk = sk_next(sk);
+try_again:
+ ;
+ } while (sk && sk->sk_family != state->family);
+
+ if (!sk && ++state->bucket < UDP_HTABLE_SIZE) {
+ sk = sk_head(&udp_hash[state->bucket]);
+ goto try_again;
+ }
+ return sk;
+}
+
+static struct sock *udp_get_idx(struct seq_file *seq, loff_t pos)
+{
+ struct sock *sk = udp_get_first(seq);
+
+ if (sk)
+ while(pos && (sk = udp_get_next(seq, sk)) != NULL)
+ --pos;
+ return pos ? NULL : sk;
+}
+
static void *udp_seq_start(struct seq_file *seq, loff_t *pos)
{
read_lock(&udp_hash_lock);
- return *pos ? udp_get_bucket(seq, pos) : (void *)1;
+ return *pos ? udp_get_idx(seq, *pos-1) : (void *)1;
}

static void *udp_seq_next(struct seq_file *seq, void *v, loff_t *pos)
{
struct sock *sk;
- struct hlist_node *node;
- struct udp_iter_state *state;
-
- if (v == (void *)1) {
- sk = udp_get_bucket(seq, pos);
- goto out;
- }

- state = seq->private;
+ if (v == (void *)1)
+ sk = udp_get_idx(seq, 0);
+ else
+ sk = udp_get_next(seq, v);

- sk = v;
- sk_for_each_continue(sk, node)
- if (sk->sk_family == state->family)
- goto out;
-
- if (++state->bucket >= UDP_HTABLE_SIZE)
- goto out;
-
- *pos = 0;
- sk = udp_get_bucket(seq, pos);
-out:
++*pos;
return sk;
}

--
Hideaki YOSHIFUJI @ USAGI Project <[email protected]>
GPG FP: 9022 65EB 1ECF 3AD1 0BDF 80D8 4807 F894 E062 0EEA

2003-09-04 07:56:18

by David Miller

[permalink] [raw]
Subject: Re: /proc/net/* read drops data

On Wed, 03 Sep 2003 16:07:33 +0900 (JST)
YOSHIFUJI Hideaki / _$B5HF#1QL@ <[email protected]> wrote:

> D: Fixing a bug that reading /proc/net/{udp,udp6} may drop some data

This fix looks good to me. Applied.

2003-09-04 16:26:18

by Ricky Beam

[permalink] [raw]
Subject: Re: /proc/net/* read drops data

On Thu, 4 Sep 2003, David S. Miller wrote:
>> D: Fixing a bug that reading /proc/net/{udp,udp6} may drop some data
>
>This fix looks good to me. Applied.

That might fix udp, but that's not the only one to be fixed. I'll compile
a list. (tcp for sure.)

--Ricky


2003-09-04 16:28:53

by Ricky Beam

[permalink] [raw]
Subject: Re: /proc/net/* read drops data

On Thu, 4 Sep 2003, David S. Miller wrote:
>> D: Fixing a bug that reading /proc/net/{udp,udp6} may drop some data
>
>This fix looks good to me. Applied.

The list:
(file) (bs=1) (bs=10000)
/proc/net/igmp 122 191
/proc/net/route 128 384
/proc/net/rt_acct 0 4096
/proc/net/rt_cache 384 512
/proc/net/tcp 1800 1950
/proc/net/udp 1024 1152

(I don't have ipv6 enabled)

--Ricky


2003-09-04 17:12:40

by YOSHIFUJI Hideaki

[permalink] [raw]
Subject: Re: /proc/net/* read drops data

In article <[email protected]> (at Thu, 4 Sep 2003 12:25:16 -0400 (EDT)), Ricky Beam <[email protected]> says:

> The list:
> (file) (bs=1) (bs=10000)
> /proc/net/igmp 122 191
> /proc/net/route 128 384
> /proc/net/rt_acct 0 4096
> /proc/net/rt_cache 384 512
> /proc/net/tcp 1800 1950
> /proc/net/udp 1024 1152

Okay, I'll seek the code.
--yoshfuji