Re: [Qemu-devel] [RFC PATCH V3 2/4] colo-compare: track connection and enqueue packet

[Jason Wang]

On 04/18/2016 07:11 PM, Zhang Chen wrote:
> In this patch we use kernel jhash table to track
> connection, and then enqueue net packet like this:
>
> + CompareState ++
> |               |
> +---------------+   +---------------+         +---------------+
> |conn list      +--->conn           +--------->conn           |
> +---------------+   +---------------+         +---------------+
> |               |     |           |             |          |
> +---------------+ +---v----+  +---v----+    +---v----+ +---v----+
>                   |primary |  |secondary    |primary | |secondary
>                   |packet  |  |packet  +    |packet  | |packet  +
>                   +--------+  +--------+    +--------+ +--------+
>                       |           |             |          |
>                   +---v----+  +---v----+    +---v----+ +---v----+
>                   |primary |  |secondary    |primary | |secondary
>                   |packet  |  |packet  +    |packet  | |packet  +
>                   +--------+  +--------+    +--------+ +--------+
>                       |           |             |          |
>                   +---v----+  +---v----+    +---v----+ +---v----+
>                   |primary |  |secondary    |primary | |secondary
>                   |packet  |  |packet  +    |packet  | |packet  +
>                   +--------+  +--------+    +--------+ +--------+
>
> Signed-off-by: Zhang Chen <address@hidden>
> Signed-off-by: Li Zhijian <address@hidden>
> Signed-off-by: Wen Congyang <address@hidden>
> ---
>  include/qemu/jhash.h |  59 ++++++++++
>  net/colo-compare.c   | 303 
> ++++++++++++++++++++++++++++++++++++++++++++++++++-
>  trace-events         |   3 +
>  3 files changed, 360 insertions(+), 5 deletions(-)
>  create mode 100644 include/qemu/jhash.h
>
> diff --git a/include/qemu/jhash.h b/include/qemu/jhash.h
> new file mode 100644
> index 0000000..8a8ff0f
> --- /dev/null
> +++ b/include/qemu/jhash.h
> @@ -0,0 +1,59 @@
> +/* jhash.h: Jenkins hash support.
> +  *
> +  * Copyright (C) 2006. Bob Jenkins (address@hidden)
> +  *
> +  * http://burtleburtle.net/bob/hash/
> +  *
> +  * These are the credits from Bob's sources:
> +  *
> +  * lookup3.c, by Bob Jenkins, May 2006, Public Domain.
> +  *
> +  * These are functions for producing 32-bit hashes for hash table lookup.
> +  * hashword(), hashlittle(), hashlittle2(), hashbig(), mix(), and final()
> +  * are externally useful functions.  Routines to test the hash are included
> +  * if SELF_TEST is defined.  You can use this free for any purpose.It's in
> +  * the public domain.  It has no warranty.
> +  *
> +  * Copyright (C) 2009-2010 Jozsef Kadlecsik (address@hidden)
> +  *
> +  * I've modified Bob's hash to be useful in the Linux kernel, and
> +  * any bugs present are my fault.
> +  * Jozsef
> +  */
> +
> +#ifndef QEMU_JHASH_H__
> +#define QEMU_JHASH_H__
> +
> +#include "qemu/bitops.h"
> +
> +/*
> + * hashtable related is copied from linux kernel jhash
> + */
> +
> +/* __jhash_mix -- mix 3 32-bit values reversibly. */
> +#define __jhash_mix(a, b, c)                \
> +{                                           \
> +    a -= c;  a ^= rol32(c, 4);  c += b;     \
> +    b -= a;  b ^= rol32(a, 6);  a += c;     \
> +    c -= b;  c ^= rol32(b, 8);  b += a;     \
> +    a -= c;  a ^= rol32(c, 16); c += b;     \
> +    b -= a;  b ^= rol32(a, 19); a += c;     \
> +    c -= b;  c ^= rol32(b, 4);  b += a;     \
> +}
> +
> +/* __jhash_final - final mixing of 3 32-bit values (a,b,c) into c */
> +#define __jhash_final(a, b, c)  \
> +{                               \
> +    c ^= b; c -= rol32(b, 14);  \
> +    a ^= c; a -= rol32(c, 11);  \
> +    b ^= a; b -= rol32(a, 25);  \
> +    c ^= b; c -= rol32(b, 16);  \
> +    a ^= c; a -= rol32(c, 4);   \
> +    b ^= a; b -= rol32(a, 14);  \
> +    c ^= b; c -= rol32(b, 24);  \
> +}
> +
> +/* An arbitrary initial parameter */
> +#define JHASH_INITVAL           0xdeadbeef
> +
> +#endif /* QEMU_JHASH_H__ */
> diff --git a/net/colo-compare.c b/net/colo-compare.c
> index c45b132..dc57eac 100644
> --- a/net/colo-compare.c
> +++ b/net/colo-compare.c
> @@ -22,12 +22,16 @@
>  #include "qemu/sockets.h"
>  #include "qapi-visit.h"
>  #include "trace.h"
> +#include "slirp/slirp.h"
> +#include "qemu/jhash.h"
> +#include "net/eth.h"
>  
>  #define TYPE_COLO_COMPARE "colo-compare"
>  #define COLO_COMPARE(obj) \
>      OBJECT_CHECK(CompareState, (obj), TYPE_COLO_COMPARE)
>  
>  #define COMPARE_READ_LEN_MAX NET_BUFSIZE
> +#define HASHTABLE_MAX_SIZE 16384
>  
>  static QTAILQ_HEAD(, CompareState) net_compares =
>         QTAILQ_HEAD_INITIALIZER(net_compares);
> @@ -39,6 +43,28 @@ typedef struct ReadState {
>      uint8_t buf[COMPARE_READ_LEN_MAX];
>  } ReadState;
>  
> +/*
> +  + CompareState ++
> +  |               |
> +  +---------------+   +---------------+         +---------------+
> +  |conn list      +--->conn           +--------->conn           |
> +  +---------------+   +---------------+         +---------------+
> +  |               |     |           |             |          |
> +  +---------------+ +---v----+  +---v----+    +---v----+ +---v----+
> +                    |primary |  |secondary    |primary | |secondary
> +                    |packet  |  |packet  +    |packet  | |packet  +
> +                    +--------+  +--------+    +--------+ +--------+
> +                        |           |             |          |
> +                    +---v----+  +---v----+    +---v----+ +---v----+
> +                    |primary |  |secondary    |primary | |secondary
> +                    |packet  |  |packet  +    |packet  | |packet  +
> +                    +--------+  +--------+    +--------+ +--------+
> +                        |           |             |          |
> +                    +---v----+  +---v----+    +---v----+ +---v----+
> +                    |primary |  |secondary    |primary | |secondary
> +                    |packet  |  |packet  +    |packet  | |packet  +
> +                    +--------+  +--------+    +--------+ +--------+
> +*/
>  typedef struct CompareState {
>      Object parent;
>  
> @@ -51,12 +77,265 @@ typedef struct CompareState {
>      QTAILQ_ENTRY(CompareState) next;
>      ReadState pri_rs;
>      ReadState sec_rs;
> +
> +    /* connection list: the connections belonged to this NIC could be found
> +     * in this list.
> +     * element type: Connection
> +     */
> +    GQueue conn_list;
> +    QemuMutex conn_list_lock; /* to protect conn_list */
> +    /* hashtable to save connection */
> +    GHashTable *connection_track_table;
> +    /* to save unprocessed_connections */
> +    GQueue unprocessed_connections;
> +    /* proxy current hash size */
> +    uint32_t hashtable_size;
>  } CompareState;
>  
>  typedef struct CompareClass {
>      ObjectClass parent_class;
>  } CompareClass;
>  
> +typedef struct Packet {
> +    void *data;
> +    union {
> +        uint8_t *network_layer;
> +        struct ip *ip;

Does this mean ipv6 is not supported?

> +    };
> +    uint8_t *transport_layer;
> +    int size;
> +    CompareState *s;
> +} Packet;
> +
> +typedef struct ConnectionKey {
> +    /* (src, dst) must be grouped, in the same way than in IP header */
> +    struct in_addr src;
> +    struct in_addr dst;
> +    uint16_t src_port;
> +    uint16_t dst_port;
> +    uint8_t ip_proto;
> +} QEMU_PACKED ConnectionKey;
> +
> +typedef struct Connection {
> +    QemuMutex list_lock;
> +    /* connection primary send queue: element type: Packet */
> +    GQueue primary_list;
> +    /* connection secondary send queue: element type: Packet */
> +    GQueue secondary_list;
> +    /* flag to enqueue unprocessed_connections */
> +    bool processing;
> +    uint8_t ip_proto;
> +} Connection;
> +
> +enum {
> +    PRIMARY_IN = 0,
> +    SECONDARY_IN,
> +};
> +
> +static void packet_destroy(void *opaque, void *user_data);
> +static int compare_chr_send(CharDriverState *out,
> +                            const uint8_t *buf,
> +                            uint32_t size);
> +
> +static uint32_t connection_key_hash(const void *opaque)
> +{
> +    const ConnectionKey *key = opaque;
> +    uint32_t a, b, c;
> +
> +    /* Jenkins hash */
> +    a = b = c = JHASH_INITVAL + sizeof(*key);
> +    a += key->src.s_addr;
> +    b += key->dst.s_addr;
> +    c += (key->src_port | key->dst_port << 16);
> +    __jhash_mix(a, b, c);
> +
> +    a += key->ip_proto;
> +    __jhash_final(a, b, c);
> +
> +    return c;
> +}
> +
> +static int connection_key_equal(const void *opaque1, const void *opaque2)
> +{
> +    return memcmp(opaque1, opaque2, sizeof(ConnectionKey)) == 0;

So why not useing ConnectionKey * consider we're sure of the type?

> +}
> +
> +/*
> + *  initialize connecon_key for packet
> + *  Return 0 on success, if return 1 the pkt will be sent later
> + */
> +static int connection_key_init(Packet *pkt, ConnectionKey *key)
> +{
> +    int network_length;
> +    uint8_t *data = pkt->data;
> +    uint16_t l3_proto;
> +    uint32_t tmp_ports;
> +    ssize_t l2hdr_len = eth_get_l2_hdr_length(data);
> +
> +    pkt->network_layer = data + ETH_HLEN;

Can the length of data be shorter than ETH_HELN?

> +    l3_proto = eth_get_l3_proto(data, l2hdr_len);
> +    if (l3_proto != ETH_P_IP) {
> +        return 1;
> +    }
> +
> +    network_length = pkt->ip->ip_hl * 4;
> +    pkt->transport_layer = pkt->network_layer + network_length;

Do we need sanity check to make sure there's no evil network_length here?

> +    if (!pkt->transport_layer) {
> +        error_report("pkt->transport_layer is valid");

invalid? And if this is caused by the bad packet it self, there's no
need for a error_report.

> +        return 1;
> +    }
> +    key->ip_proto = pkt->ip->ip_p;
> +    key->src = pkt->ip->ip_src;
> +    key->dst = pkt->ip->ip_dst;
> +
> +    switch (key->ip_proto) {
> +    case IPPROTO_TCP:
> +    case IPPROTO_UDP:
> +    case IPPROTO_DCCP:
> +    case IPPROTO_ESP:
> +    case IPPROTO_SCTP:
> +    case IPPROTO_UDPLITE:
> +        tmp_ports = *(uint32_t *)(pkt->transport_layer);
> +        key->src_port = ntohs(tmp_ports & 0xffff);
> +        key->dst_port = ntohs(tmp_ports >> 16);
> +        break;
> +    case IPPROTO_AH:
> +        tmp_ports = *(uint32_t *)(pkt->transport_layer + 4);
> +        key->src_port = ntohs(tmp_ports & 0xffff);
> +        key->dst_port = ntohs(tmp_ports >> 16);
> +        break;
> +    default:
> +        key->src_port = 0;
> +        key->dst_port = 0;
> +        break;
> +    }
> +
> +    return 0;
> +}
> +
> +static Connection *connection_new(ConnectionKey *key)
> +{
> +    Connection *conn = g_slice_new(Connection);
> +
> +    qemu_mutex_init(&conn->list_lock);
> +    conn->ip_proto = key->ip_proto;
> +    conn->processing = false;
> +    g_queue_init(&conn->primary_list);
> +    g_queue_init(&conn->secondary_list);
> +
> +    return conn;
> +}
> +
> +/*
> + * Clear hashtable, stop this hash growing really huge
> + */
> +static void connection_hashtable_reset(CompareState *s)
> +{
> +    s->hashtable_size = 0;
> +    g_hash_table_remove_all(s->connection_track_table);
> +}
> +
> +/* if not found, create a new connection and add to hash table */
> +static Connection *connection_get(CompareState *s, ConnectionKey *key)
> +{
> +    /* FIXME: protect connection_track_table */
> +    Connection *conn = g_hash_table_lookup(s->connection_track_table, key);
> +
> +    if (conn == NULL) {
> +        ConnectionKey *new_key = g_memdup(key, sizeof(*key));
> +
> +        conn = connection_new(key);
> +
> +        s->hashtable_size++;
> +        if (s->hashtable_size > HASHTABLE_MAX_SIZE) {
> +            error_report("colo proxy connection hashtable full, clear it");
> +            connection_hashtable_reset(s);
> +            /* TODO:clear conn_list */
> +        }
> +
> +        g_hash_table_insert(s->connection_track_table, new_key, conn);
> +    }
> +
> +     return conn;
> +}
> +
> +static void connection_destroy(void *opaque)
> +{
> +    Connection *conn = opaque;
> +
> +    qemu_mutex_lock(&conn->list_lock);

Like I said in previous patch, if you do all the processing in colo
compare thread, you can avoid almost all synchronization (e.g mutex).

> +    g_queue_foreach(&conn->primary_list, packet_destroy, NULL);
> +    g_queue_free(&conn->primary_list);
> +    g_queue_foreach(&conn->secondary_list, packet_destroy, NULL);
> +    g_queue_free(&conn->secondary_list);
> +    qemu_mutex_unlock(&conn->list_lock);
> +    qemu_mutex_destroy(&conn->list_lock);
> +    g_slice_free(Connection, conn);
> +}
> +
> +static Packet *packet_new(CompareState *s, const void *data,
> +                              int size, ConnectionKey *key)
> +{
> +    Packet *pkt = g_slice_new(Packet);
> +
> +    pkt->data = g_memdup(data, size);
> +    pkt->size = size;
> +    pkt->s = s;
> +
> +    if (connection_key_init(pkt, key)) {
> +        packet_destroy(pkt, NULL);
> +        pkt = NULL;
> +    }

Can we do connection_key_init() first, this can avoid packet_desctory()
if it fails.

> +
> +    return pkt;
> +}
> +
> +/*
> + * Return 0 on success, if return -1 means the pkt
> + * is unsupported(arp and ipv6) and will be sent later
> + */
> +static int packet_enqueue(CompareState *s, int mode)
> +{
> +    ConnectionKey key = {{ 0 } };
> +    Packet *pkt = NULL;
> +    Connection *conn;
> +
> +    if (mode == PRIMARY_IN) {
> +        pkt = packet_new(s, s->pri_rs.buf, s->pri_rs.packet_len, &key);
> +    } else {
> +        pkt = packet_new(s, s->sec_rs.buf, s->sec_rs.packet_len, &key);
> +    }
> +    if (!pkt) {
> +        return -1;
> +    }
> +
> +    conn = connection_get(s, &key);
> +    if (!conn->processing) {
> +        qemu_mutex_lock(&s->conn_list_lock);
> +        g_queue_push_tail(&s->conn_list, conn);
> +        qemu_mutex_unlock(&s->conn_list_lock);
> +        conn->processing = true;
> +    }
> +
> +    qemu_mutex_lock(&conn->list_lock);
> +    if (mode == PRIMARY_IN) {
> +        g_queue_push_tail(&conn->primary_list, pkt);
> +    } else {
> +        g_queue_push_tail(&conn->secondary_list, pkt);
> +    }
> +    qemu_mutex_unlock(&conn->list_lock);
> +
> +    return 0;
> +}
> +
> +static void packet_destroy(void *opaque, void *user_data)
> +{
> +    Packet *pkt = opaque;
> +
> +    g_free(pkt->data);
> +    g_slice_free(Packet, pkt);
> +}
> +
>  static int compare_chr_send(CharDriverState *out,
>                              const uint8_t *buf,
>                              uint32_t size)
> @@ -158,8 +437,10 @@ static void compare_pri_chr_in(void *opaque, const 
> uint8_t *buf, int size)
>  
>      ret = compare_chr_fill_rstate(&s->pri_rs, buf, size);
>      if (ret == 1) {
> -        /* FIXME: enqueue to primary packet list */
> -        compare_chr_send(s->chr_out, s->pri_rs.buf, s->pri_rs.packet_len);
> +        if (packet_enqueue(s, PRIMARY_IN)) {
> +            trace_colo_compare_main("primary: unsupported packet in");
> +            compare_chr_send(s->chr_out, s->pri_rs.buf, 
> s->pri_rs.packet_len);

Looks like if a packet was not recognized by connection_key_init(), it
will be sent directly without comparing it with the packet sent from
secondary? Is this expected?

> +        }
>      } else if (ret == -1) {
>          qemu_chr_add_handlers(s->chr_pri_in, NULL, NULL, NULL, NULL);
>      }
> @@ -176,9 +457,11 @@ static void compare_sec_chr_in(void *opaque, const 
> uint8_t *buf, int size)
>  
>      ret = compare_chr_fill_rstate(&s->sec_rs, buf, size);
>      if (ret == 1) {
> -        /* TODO: enqueue to secondary packet list*/
> -        /* should we send sec arp pkt? */
> -        compare_chr_send(s->chr_out, s->sec_rs.buf, s->sec_rs.packet_len);
> +        if (packet_enqueue(s, SECONDARY_IN)) {
> +            trace_colo_compare_main("secondary: unsupported packet in");
> +            /* should we send sec arp pkt? */
> +            compare_chr_send(s->chr_out, s->sec_rs.buf, 
> s->sec_rs.packet_len);
> +        }
>      } else if (ret == -1) {
>          qemu_chr_add_handlers(s->chr_sec_in, NULL, NULL, NULL, NULL);
>      }
> @@ -280,6 +563,15 @@ static void colo_compare_complete(UserCreatable *uc, 
> Error **errp)
>      qemu_chr_fe_claim_no_fail(s->chr_out);
>      QTAILQ_INSERT_TAIL(&net_compares, s, next);
>  
> +    g_queue_init(&s->conn_list);
> +    qemu_mutex_init(&s->conn_list_lock);
> +    s->hashtable_size = 0;
> +
> +    s->connection_track_table = g_hash_table_new_full(connection_key_hash,
> +                                                      connection_key_equal,
> +                                                      g_free,
> +                                                      connection_destroy);
> +
>      return;
>  }
>  
> @@ -314,6 +606,7 @@ static void colo_compare_class_finalize(ObjectClass *oc, 
> void *data)
>      if (!QTAILQ_EMPTY(&net_compares)) {
>          QTAILQ_REMOVE(&net_compares, s, next);
>      }
> +    qemu_mutex_destroy(&s->conn_list_lock);
>  }
>  
>  static void colo_compare_init(Object *obj)
> diff --git a/trace-events b/trace-events
> index ca7211b..8862288 100644
> --- a/trace-events
> +++ b/trace-events
> @@ -1916,3 +1916,6 @@ aspeed_vic_update_fiq(int flags) "Raising FIQ: %d"
>  aspeed_vic_update_irq(int flags) "Raising IRQ: %d"
>  aspeed_vic_read(uint64_t offset, unsigned size, uint32_t value) "From 0x%" 
> PRIx64 " of size %u: 0x%" PRIx32
>  aspeed_vic_write(uint64_t offset, unsigned size, uint32_t data) "To 0x%" 
> PRIx64 " of size %u: 0x%" PRIx32
> +
> +# net/colo-compare.c
> +colo_compare_main(const char *chr) "chr: %s"