OpenVZ Forum


Home » Mailing lists » Users » Networkproblem: VE stops sending ACKs
Networkproblem: VE stops sending ACKs [message #8686] Sat, 02 December 2006 00:43 Go to next message
kaymes
Messages: 3
Registered: December 2006
Junior Member
Hi!
I've got a strange networkingproblem around here - the VE suddenly stops to
send ACKs.

Setup is a VE with its own ip and proxyarp/iptables running on HN. Sometimes a
tcp/ip-connection simply hangs and does not recover at all. I did look into
it with ethereal and discovered the following situation:

The VE was downloading a larger file via http on a fast connection. Apparently
a packet got lost, so the VE started to send dupACKs. Of course the sender
kept sending packets, because the dupACKs hadn't arrived jet. So far so good.
However, after 155 (154 in another case) dupACKs, the VE simply stopped
sending any ACKs whatsoever. The sender kept retransmitting the missing
packet, but the VE simply stayed quiet.

Any idea about what's going on here and how to solve this problem?

Cheers,
Konstantin
Re: Networkproblem: VE stops sending ACKs [message #8693 is a reply to message #8686] Sat, 02 December 2006 13:55 Go to previous messageGo to next message
kaymes
Messages: 3
Registered: December 2006
Junior Member
On Saturday 02 December 2006 01:43, Konstantin Seiler wrote:
> I've got a strange networkingproblem around here - the VE suddenly stops to
> send ACKs.

Hi!
I finally found something related to that problem. It seems to be a problem
with TCPRCVBUF in the VEs config. For a check I multiplied the values by 10
and the download went fine.
(By the way, what are considered proper values - the defaults caused the
trouble)

So I interpret the situation as a Bug of OpenVZ: The VE doesn't take TCPRCVBUF
into account when setting the TCP-Window-Size and the propagated window is
too big for the available buffer.
If a packet gets lost now, the connection is doomed to die: The sender keeps
sending packets and eventually the buffer is full. When the retransmission of
the lost packet finally arrives, it is discarded because of the full buffer
and the connection is in a deadlock.
The retransmission can't be accepted because of the full buffer and the buffer
can't be emptied, because a packet is missing.

Cheers,
Konstantin
Re: Networkproblem: VE stops sending ACKs [message #9018 is a reply to message #8693] Wed, 13 December 2006 12:54 Go to previous message
dev is currently offline  dev
Messages: 1693
Registered: September 2005
Location: Moscow
Senior Member

Konstantin,

can you please check the patch attached whether it helps?

Thanks,
Kirill

> On Saturday 02 December 2006 01:43, Konstantin Seiler wrote:
>
>>I've got a strange networkingproblem around here - the VE suddenly stops to
>>send ACKs.
>
>
> Hi!
> I finally found something related to that problem. It seems to be a problem
> with TCPRCVBUF in the VEs config. For a check I multiplied the values by 10
> and the download went fine.
> (By the way, what are considered proper values - the defaults caused the
> trouble)
>
> So I interpret the situation as a Bug of OpenVZ: The VE doesn't take TCPRCVBUF
> into account when setting the TCP-Window-Size and the propagated window is
> too big for the available buffer.
> If a packet gets lost now, the connection is doomed to die: The sender keeps
> sending packets and eventually the buffer is full. When the retransmission of
> the lost packet finally arrives, it is discarded because of the full buffer
> and the connection is in a deadlock.
> The retransmission can't be accepted because of the full buffer and the buffer
> can't be emptied, because a packet is missing.
>
> Cheers,
> Konstantin
diff --git a/include/net/tcp.h b/include/net/tcp.h
index 0ff49a5..7e8f200 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -1815,8 +1815,9 @@ static inline int tcp_win_from_space(int
/* Note: caller must be prepared to deal with negative returns */
static inline int tcp_space(const struct sock *sk)
{
- return tcp_win_from_space(sk->sk_rcvbuf -
- atomic_read(&sk->sk_rmem_alloc));
+ int ub_tcp_rcvbuf = (int) sock_bc(sk)->ub_tcp_rcvbuf;
+ return tcp_win_from_space(min(sk->sk_rcvbuf, ub_tcp_rcvbuf)
+ - atomic_read(&sk->sk_rmem_alloc));
}

static inline int tcp_full_space(const struct sock *sk)
diff --git a/include/ub/beancounter.h b/include/ub/beancounter.h
index 3d87afa..fc236e8 100644
--- a/include/ub/beancounter.h
+++ b/include/ub/beancounter.h
@@ -144,6 +144,8 @@ struct sock_private {
unsigned long ubp_rmem_thres;
unsigned long ubp_wmem_pressure;
unsigned long ubp_maxadvmss;
+ /* Total size of all advertised receive windows for all tcp sockets */
+ unsigned long ubp_rcv_wnd;
unsigned long ubp_rmem_pressure;
#define UB_RMEM_EXPAND 0
#define UB_RMEM_KEEP 1
@@ -177,6 +179,7 @@ #define ub_held_pages ppriv.ubp_held_pa
struct sock_private spriv;
#define ub_rmem_thres spriv.ubp_rmem_thres
#define ub_maxadvmss spriv.ubp_maxadvmss
+#define ub_rcv_wnd spriv.ubp_rcv_wnd
#define ub_rmem_pressure spriv.ubp_rmem_pressure
#define ub_wmem_pressure spriv.ubp_wmem_pressure
#define ub_tcp_sk_list spriv.ubp_tcp_socks
diff --git a/include/ub/ub_sk.h b/include/ub/ub_sk.h
index e65c9ed..02d0137 100644
--- a/include/ub/ub_sk.h
+++ b/include/ub/ub_sk.h
@@ -34,6 +34,8 @@ struct sock_beancounter {
*/
unsigned long poll_reserv;
unsigned long forw_space;
+ unsigned long ub_tcp_rcvbuf;
+ unsigned long ub_rcv_wnd_old;
/* fields below are protected by bc spinlock */
unsigned long ub_waitspc; /* space waiting for */
unsigned long ub_wcharged;
diff --git a/kernel/ub/ub_net.c b/kernel/ub/ub_net.c
index 74d651a..afee710 100644
--- a/kernel/ub/ub_net.c
+++ b/kernel/ub/ub_net.c
@@ -420,6 +420,7 @@ static int __sock_charge(struct sock *sk

added_reserv = 0;
added_forw = 0;
+ skbc->ub_rcv_wnd_old = 0;
if (res == UB_NUMTCPSOCK) {
added_reserv = skb_charge_size(MAX_TCP_HEADER +
1500 - sizeof(struct iphdr) -
@@ -439,6 +440,7 @@ static int __sock_charge(struct sock *sk
added_forw = 0;
}
skbc->forw_space = added_forw;
+ skbc->ub_tcp_rcvbuf = added_forw + SK_STREAM_MEM_QUANTUM;
}
spin_unlock_irqrestore(&ub->ub_lock, flags);

@@ -528,6 +530,7 @@ void ub_sock_uncharge(struct sock *sk)
skbc->ub_wcharged, skbc->ub, skbc->ub->ub_uid);
skbc->poll_reserv = 0;
skbc->forw_space = 0;
+ ub->ub_rcv_wnd -= is_tcp_sock ? tcp_sk(sk)->rcv_wnd : 0;
spin_unlock_irqrestore(&ub->ub_lock, flags);

uncharge_beancounter_notop(skbc->ub,
@@ -768,6 +771,44 @@ static void ub_sockrcvbuf_uncharge(struc
* UB_TCPRCVBUF
*/

+/*
+ * UBC TCP window management mechanism.
+ * Every socket may consume no more than sock_quantum.
+ * sock_quantum depends on space available and ub_parms[UB_NUMTCPSOCK].held.
+ */
+static void ub_sock_tcp_update_rcvbuf(struct user_beancounter *ub,
+ struct sock *sk)
+{
+ unsigned long allowed;
+ unsigned long reserved;
+ unsigned long available;
+ unsigned long sock_quantum;
+ struct tcp_opt *tp = tcp_sk(sk);
+ struct sock_beancounter *skbc;
+ skbc = sock_bc(sk);
+
+ if( ub->ub_parms[UB_NUMTCPSOCK].limit * ub->ub_maxadvmss
+ > ub->ub_parms[UB_TCPRCVBUF].limit) {
+ /* this is defenitly shouldn't happend */
+ return;
+ }
+ allowed = ub->ub_parms[UB_TCPRCVBUF].barrier;
+ ub->ub_rcv_wnd += (tp->rcv_wnd - skbc->ub_rcv_wnd_old);
+ skbc->ub_rcv_wnd_old = tp->rcv_wnd;
+ reserved = ub->ub_parms[UB_TCPRCVBUF].held + ub->ub_rcv_wnd;
+ available = (allowed < reserved)?
+ 0:allowed - reserved;
+ sock_quantum = max(allowed / ub->ub_parms[UB_NUMTCPSOCK].held,
+ ub->ub_maxadvmss);
+ if ( skbc->ub_tcp_rcvbuf > sock_quantum) {
+ skbc->ub_tcp_rcvbuf = sock_quantum;
+ } else {
+ skbc->ub_tcp_rcvbuf += min(sock_quantum - skbc->ub_tcp_rcvbuf,
+ available);
+ }
+
+}
+
int ub_sock_tcp_chargerecv(struct sock *sk, struct sk_buff *skb,
enum ub_severity strict)
{
@@ -804,6 +845,7 @@ int ub_sock_tcp_chargerecv(struct sock *
retval = 0;
for (ub = sock_bc(sk)->ub; ub->parent != NULL; ub = ub->parent);
spin_lock_irqsave(&ub->ub_lock, flags);
+ ub_sock_tcp_update_rcvbuf(ub, sk);
ub->ub_parms[UB_TCPRCVBUF].held += chargesize;
if (ub->ub_parms[UB_TCPRCVBUF].held >
ub->ub_parms[UB_TCPRCVBUF].barrier &&
Previous Topic: CfP Workshops on Virtualization/XEN in HPC Cluster and Grid Computing Environments (XHPC'07, VHPC'07
Next Topic: Load Sharing/Balancing and Standby Redundacy among VMs
Goto Forum:
  


Current Time: Tue Feb 27 15:05:48 GMT 2024

Total time taken to generate the page: 0.02611 seconds