|
From: | Jason Wang |
Subject: | Re: [Qemu-devel] [PATCH v2 2/3] filter-rewriter: fix memory leak for connection in connection_track_table |
Date: | Mon, 27 Feb 2017 11:40:11 +0800 |
User-agent: | Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.7.0 |
On 2017年02月27日 11:11, Hailiang Zhang wrote:
On 2017/2/23 12:16, Jason Wang wrote:On 2017年02月22日 16:51, Hailiang Zhang wrote:On 2017/2/22 16:45, Hailiang Zhang wrote:On 2017/2/22 16:07, Jason Wang wrote:On 2017年02月22日 11:46, zhanghailiang wrote:After a net connection is closed, we didn't clear its releated resources in connection_track_table, which will lead to memory leak.Not a real leak but would lead reset of hash table if too many closed connections.Yes, you are right, there will be lots of stale connection data in hash table if we don't remove it while it is been closed. WhichOk, so let's come up with a better title of the patch.OK.Let't track the state of net connection, if it is closed, its relatedresources will be cleared up.The issue is the state were tracked partially, do we need a full statemachine here?Not, IMHO, we only care about the last state of it, because, we will do nothing even if we track the intermedial states.Well, you care at least syn state too. Without a complete state machine, it's very hard to track even partial state I believe. And you will fail to track some state transition for sure which makes the code fragile.Agree, but here things are a little different. There are some extreme casesthat we may can't track the complete process of closing connection.For example (I have explained that in the bellow, it seems that you didn'tgot it ;) ). If VM is running before we want to make it goes into COLO FT state,there maybe some connections exist already, in extreme case, VM is going into COLO state while some connections are in half closing state, we can only trackthe bellow half closing state in filter-rewriter and colo compare object.Signed-off-by: zhanghailiang <address@hidden> --- net/colo.h | 4 +++ net/filter-rewriter.c | 70 +++++++++++++++++++++++++++++++++++++++++++++------ 2 files changed, 67 insertions(+), 7 deletions(-) diff --git a/net/colo.h b/net/colo.h index 7c524f3..cd9027f 100644 --- a/net/colo.h +++ b/net/colo.h @@ -18,6 +18,7 @@ #include "slirp/slirp.h" #include "qemu/jhash.h" #include "qemu/timer.h" +#include "slirp/tcp.h" #define HASHTABLE_MAX_SIZE 16384 @@ -69,6 +70,9 @@ typedef struct Connection { * run once in independent tcp connection */ int syn_flag; + + int tcp_state; /* TCP FSM state */ + tcp_seq fin_ack_seq; /* the seq of 'fin=1,ack=1' */ } Connection; uint32_t connection_key_hash(const void *opaque); diff --git a/net/filter-rewriter.c b/net/filter-rewriter.c index c4ab91c..7e7ec35 100644 --- a/net/filter-rewriter.c +++ b/net/filter-rewriter.c @@ -60,9 +60,9 @@ static int is_tcp_packet(Packet *pkt) } /* handle tcp packet from primary guest */ -static int handle_primary_tcp_pkt(NetFilterState *nf, +static int handle_primary_tcp_pkt(RewriterState *rf, Connection *conn, - Packet *pkt) + Packet *pkt, ConnectionKey *key) { struct tcphdr *tcp_pkt; @@ -97,15 +97,45 @@ static int handle_primary_tcp_pkt(NetFilterState *nf, tcp_pkt->th_ack = htonl(ntohl(tcp_pkt->th_ack) + conn->offset);net_checksum_calculate((uint8_t *)pkt->data, pkt->size);+ /* + * Case 1: + * The *server* side of this connect is VM, *client* tries to close + * the connection. + * + * We got 'ack=1' packets from client side, it acks 'fin=1, ack=1' + * packet from server side. From this point, we can ensure that there + * will be no packets in the connection, except that, some errors + * happen between the path of 'filter object' and vNIC, if this rare+ * case really happen, we can still create a new connection,+ * So it is safe to remove the connection from connection_track_table. + * + */ + if ((conn->tcp_state == TCPS_LAST_ACK) && + (ntohl(tcp_pkt->th_ack) == (conn->fin_ack_seq + 1))) { + fprintf(stderr, "Remove conn "Can this even compile?Oops, i forgot to remove it, will remove it in next version.+ g_hash_table_remove(rf->connection_track_table, key); + } + } + /* + * Case 2: + * The *server* side of this connect is VM, *server* tries to close + * the connection. + * + * We got 'fin=1, ack=1' packet from client side, we need to + * record the seq of 'fin=1, ack=1' packet. + */ + if ((tcp_pkt->th_flags & (TH_ACK | TH_FIN)) == (TH_ACK | TH_FIN)) { + conn->fin_ack_seq = htonl(tcp_pkt->th_seq); + conn->tcp_state = TCPS_LAST_ACK; } return 0; } /* handle tcp packet from secondary guest */ -static int handle_secondary_tcp_pkt(NetFilterState *nf, +static int handle_secondary_tcp_pkt(RewriterState *rf, Connection *conn, - Packet *pkt)+ Packet *pkt, ConnectionKey *key){ struct tcphdr *tcp_pkt; @@ -133,8 +163,34 @@ static int handle_secondary_tcp_pkt(NetFilterState *nf, tcp_pkt->th_seq = htonl(ntohl(tcp_pkt->th_seq) - conn->offset);net_checksum_calculate((uint8_t *)pkt->data, pkt->size);+ /* + * Case 2: + * The *server* side of this connect is VM, *server* tries to close + * the connection. + * + * We got 'ack=1' packets from server side, it acks 'fin=1, ack=1' + * packet from client side. Like Case 1, there should be no packets + * in the connection from now know, But the difference here is + * if the packet is lost, We will get the resent 'fin=1,ack=1' packet. + * TODO: Fix above case. + */ + if ((conn->tcp_state == TCPS_LAST_ACK) && + (ntohl(tcp_pkt->th_ack) == (conn->fin_ack_seq + 1))) { + g_hash_table_remove(rf->connection_track_table, key); + } + } + /* + * Case 1: + * The *server* side of this connect is VM, *client* tries to close + * the connection. + * + * We got 'fin=1, ack=1' packet from server side, we need to + * record the seq of 'fin=1, ack=1' packet. + */ + if ((tcp_pkt->th_flags & (TH_ACK | TH_FIN)) == (TH_ACK | TH_FIN)) { + conn->fin_ack_seq = ntohl(tcp_pkt->th_seq); + conn->tcp_state = TCPS_LAST_ACK;I thought the tcp_state should store the state of TCP from the view ofsecondary VM? So TCPS_LAST_ACK is wrong and bring lots of confusion. Andthe handle of active close needs more states here. E.g if connection isin FIN_WAIT_2, the connection is only half closed, remote peer can still send packet to us unless we receive a FIN.Yes, i know what you mean, actually, here, we try to only track the last two steps for closing a connection, that is 'fin=1,ack=1,seq=2,ack=u+1'^ 'FIN=1,ACK=1,seq=w,ack=u+1'and 'ack=1,seq=u+1,ack=w+1', because if we get a 'fin=1,ack=1', we can^ 'ACK=1,seq=u+1,ack=w+1' ^ 'FIN=1,ACK=1'ensure that the 'fin=1,seq=u' packet has been posted.^ 'FIN=1,seq=u'That's just the case I'm saying, the transition above is in fact: secondary(ESTABLISHED) secondary(FIN_WAIT_1): -> FIN,seq=w,ack=u+1 -> :remote secondary(FIN_WAIT_2): <- seq=u+1,ack=w+1 <- :remote So we are in fact in FIN_WAIT_2, which means the connection is only half closed, but your patch will treat this as fully closed connection and will remove the connection from the hashtable.Er, here we track the last two states 'FIN=1, ACK=1' and 'ACK=1' ( which asks the 'FIN=1,ACK=1' packet, We will remove the connection while got the 'ACK=1'packet, so is it enough ?
But the connection is not closed in fact, no? It's legal for remote to continue sending tons of packet to us even after this.
What's more I don't think we can decide passive or active close by: + if ((tcp_pkt->th_flags & (TH_ACK | TH_FIN)) == (TH_ACK | TH_FIN)) { Since both cases will send FIN,ACK for sure.I didn't quite understand, here we have tracked the closing request no matter it is from the server side (passive close ?) or client side ( active close ?).You can refer to the comment in codes, 'Case 1' and 'Case 2' comments.
I think you need differ them since passive close is much simpler, and it seems that your code may work only in this case.
Here, it seems that we can't track one case which both sides send the closing requests at the same time, in that case, there are only 'FIN=1' and 'ACK=1'packets.
Yes, RFC allows this. Thanks
Thanks. HailiangAnother reason is we may can't track the 'fin=1,seq=u' packet while we start COLO while one connection is closing, which the 'fin=1,seq=u' packet has been posted. Actually, here, if we start COLO while one connection is closing, which the'fin=1,ack=1' has been posted, we can only track 'ack=1' packet. In this^ 'FIN=1,ACK=1' Sorry for the typo. :)case, the connection will be left in hash table for ever though it is harmless. Any ideas for this case ?Sorry I don't follow the question.For the above codes question, i'd like to change tcp_state to tap_closing_wait, is it OK ?You mean "tcp_closing_wait". I think we need first figure out if we can track the state correctly first. ThanksThanks. HailiangThanks} - return 0; } @@ -178,7 +234,7 @@ static ssize_t colo_rewriter_receive_iov(NetFilterState *nf, if (sender == nf->netdev) { /* NET_FILTER_DIRECTION_TX */ - if (!handle_primary_tcp_pkt(nf, conn, pkt)) { + if (!handle_primary_tcp_pkt(s, conn, pkt, &key)) { qemu_net_queue_send(s->incoming_queue, sender, 0, (const uint8_t *)pkt->data, pkt->size, NULL); packet_destroy(pkt, NULL); @@ -191,7 +247,7 @@ static ssize_t colo_rewriter_receive_iov(NetFilterState *nf, } } else { /* NET_FILTER_DIRECTION_RX */ - if (!handle_secondary_tcp_pkt(nf, conn, pkt)) { + if (!handle_secondary_tcp_pkt(s, conn, pkt, &key)) { qemu_net_queue_send(s->incoming_queue, sender, 0, (const uint8_t *)pkt->data, pkt->size, NULL); packet_destroy(pkt, NULL);..
[Prev in Thread] | Current Thread | [Next in Thread] |