|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] [PATCH net-next v1 6/8] xen-netback: add an implementation of toeplitz hashing...
On Fri, Feb 12, 2016 at 11:13 AM, Paul Durrant <paul.durrant@xxxxxxxxxx> wrote:
> ...for receive-side packets.
>
> My recent patch to include/xen/interface/io/netif.h defines a set of
> control messages that can be used by a VM frontend driver to configure
> toeplitz hashing of receive-side packets and consequent steering of those
> packets to particular queues.
>
> This patch introduces an implementation of toeplitz hashing and into
> xen-netback and allows it to be configured using the new control messages.
>
> Signed-off-by: Paul Durrant <paul.durrant@xxxxxxxxxx>
> Cc: Ian Campbell <ian.campbell@xxxxxxxxxx>
> Cc: Wei Liu <wei.liu2@xxxxxxxxxx>
> ---
> drivers/net/xen-netback/common.h | 13 ++++
> drivers/net/xen-netback/interface.c | 149
> ++++++++++++++++++++++++++++++++++++
> drivers/net/xen-netback/netback.c | 128 ++++++++++++++++++++++++++++++-
> 3 files changed, 287 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/net/xen-netback/common.h
> b/drivers/net/xen-netback/common.h
> index 093a12a..6687702 100644
> --- a/drivers/net/xen-netback/common.h
> +++ b/drivers/net/xen-netback/common.h
> @@ -220,6 +220,12 @@ struct xenvif_mcast_addr {
>
> #define XEN_NETBK_MCAST_MAX 64
>
> +#define XEN_NETBK_MAX_TOEPLITZ_KEY_SIZE 40
> +
> +#define XEN_NETBK_MAX_TOEPLITZ_MAPPING_ORDER 7
> +#define XEN_NETBK_MAX_TOEPLITZ_MAPPING_SIZE \
> + BIT(XEN_NETBK_MAX_TOEPLITZ_MAPPING_ORDER)
> +
> struct xenvif {
> /* Unique identifier for this interface. */
> domid_t domid;
> @@ -251,6 +257,13 @@ struct xenvif {
> unsigned int num_queues; /* active queues, resource allocated */
> unsigned int stalled_queues;
>
> + struct {
> + u32 flags;
> + u8 key[XEN_NETBK_MAX_TOEPLITZ_KEY_SIZE];
> + u32 mapping[XEN_NETBK_MAX_TOEPLITZ_MAPPING_SIZE];
> + unsigned int order;
> + } toeplitz;
> +
> struct xenbus_watch credit_watch;
> struct xenbus_watch mcast_ctrl_watch;
>
> diff --git a/drivers/net/xen-netback/interface.c
> b/drivers/net/xen-netback/interface.c
> index 1850ebb..230afde 100644
> --- a/drivers/net/xen-netback/interface.c
> +++ b/drivers/net/xen-netback/interface.c
> @@ -1,3 +1,4 @@
> +
> /*
> * Network-device interface management.
> *
> @@ -151,6 +152,153 @@ void xenvif_wake_queue(struct xenvif_queue *queue)
> netif_tx_wake_queue(netdev_get_tx_queue(dev, id));
> }
>
> +static u32 toeplitz_hash(const u8 *k, unsigned int klen,
> + const u8 *d, unsigned int dlen)
This should be a common library function, probably in lib directory.
> +{
> + unsigned int di, ki;
> + u64 prefix = 0;
> + u64 hash = 0;
> +
> + /* Pre-load prefix with the first 8 bytes of the key */
> + for (ki = 0; ki < 8; ki++) {
> + prefix <<= 8;
> + prefix |= (ki < klen) ? k[ki] : 0;
> + }
> +
> + for (di = 0; di < dlen; di++) {
> + u8 byte = d[di];
> + unsigned int bit;
> +
> + for (bit = 0x80; bit != 0; bit >>= 1) {
> + if (byte & bit)
> + hash ^= prefix;
> + prefix <<= 1;
> + }
> +
> + /* prefix has now been left-shifted by 8, so OR in
> + * the next byte.
> + */
> + prefix |= (ki < klen) ? k[ki] : 0;
> + ki++;
> + }
> +
> + /* The valid part of the hash is in the upper 32 bits. */
> + return hash >> 32;
> +}
> +
> +static void xenvif_set_toeplitz_hash(struct xenvif *vif, struct sk_buff *skb)
> +{
> + struct flow_keys flow;
> + u32 hash = 0;
> + enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
> + const u8 *key = vif->toeplitz.key;
> + u32 flags = vif->toeplitz.flags;
> + const unsigned int len = XEN_NETBK_MAX_TOEPLITZ_KEY_SIZE;
> + bool has_tcp_hdr;
> +
> + /* Quick rejection test: If the network protocol doesn't
> + * correspond to any enabled hash type then there's no point
> + * in parsing the packet header.
> + */
> + switch (skb->protocol) {
> + case htons(ETH_P_IP):
> + if (flags & (XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV4_TCP |
> + XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV4))
> + break;
> +
> + goto done;
> +
> + case htons(ETH_P_IPV6):
> + if (flags & (XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV6_TCP |
> + XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV6))
> + break;
> +
> + goto done;
> +
> + default:
> + goto done;
> + }
> +
> + memset(&flow, 0, sizeof(flow));
> + if (!skb_flow_dissect_flow_keys(skb, &flow, 0))
Flow dissector will parse into encapsulations to find IP addresses.
This may or may not be what you want (we'd have to look at NDIS spec).
> + goto done;
> +
> + has_tcp_hdr = (flow.basic.ip_proto == IPPROTO_TCP) &&
> + !(flow.control.flags & FLOW_DIS_IS_FRAGMENT);
> +
> + switch (skb->protocol) {
> + case htons(ETH_P_IP):
> + if (has_tcp_hdr &&
> + (flags & XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV4_TCP)) {
> + u8 data[12];
> +
> + memcpy(&data[0], &flow.addrs.v4addrs.src, 4);
> + memcpy(&data[4], &flow.addrs.v4addrs.dst, 4);
> + memcpy(&data[8], &flow.ports.src, 2);
> + memcpy(&data[10], &flow.ports.dst, 2);
> +
> + hash = toeplitz_hash(key, len,
> + data, sizeof(data));
> + type = PKT_HASH_TYPE_L4;
> + } else if (flags & XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV4) {
> + u8 data[8];
> +
> + memcpy(&data[0], &flow.addrs.v4addrs.src, 4);
> + memcpy(&data[4], &flow.addrs.v4addrs.dst, 4);
> +
> + hash = toeplitz_hash(key, len,
> + data, sizeof(data));
> + type = PKT_HASH_TYPE_L3;
> + }
> +
> + break;
> +
> + case htons(ETH_P_IPV6):
> + if (has_tcp_hdr &&
> + (flags & XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV6_TCP)) {
> + u8 data[36];
> +
> + memcpy(&data[0], &flow.addrs.v6addrs.src, 16);
> + memcpy(&data[16], &flow.addrs.v6addrs.dst, 16);
> + memcpy(&data[32], &flow.ports.src, 2);
> + memcpy(&data[34], &flow.ports.dst, 2);
> +
> + hash = toeplitz_hash(key, len,
> + data, sizeof(data));
> + type = PKT_HASH_TYPE_L4;
> + } else if (flags & XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV6) {
> + u8 data[32];
> +
> + memcpy(&data[0], &flow.addrs.v6addrs.src, 16);
> + memcpy(&data[16], &flow.addrs.v6addrs.dst, 16);
> +
> + hash = toeplitz_hash(key, len,
> + data, sizeof(data));
> + type = PKT_HASH_TYPE_L3;
> + }
> +
> + break;
> + }
> +
> +done:
> + skb_set_hash(skb, hash, type);
Is this necessary, it is potentially overwriting a valid L4 hash that
might be used later in sometling like RFS? Why not just return the
hash value.
> +}
> +
> +static u16 xenvif_select_queue(struct net_device *dev, struct sk_buff *skb,
> + void *accel_priv,
> + select_queue_fallback_t fallback)
> +{
> + struct xenvif *vif = netdev_priv(dev);
> + unsigned int mask = (1u << vif->toeplitz.order) - 1;
> +
> + if (vif->toeplitz.flags == 0)
> + return fallback(dev, skb) % dev->real_num_tx_queues;
> +
> + xenvif_set_toeplitz_hash(vif, skb);
> +
> + return vif->toeplitz.mapping[skb_get_hash_raw(skb) & mask];
> +}
> +
> static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
> {
> struct xenvif *vif = netdev_priv(dev);
> @@ -395,6 +543,7 @@ static const struct ethtool_ops xenvif_ethtool_ops = {
> };
>
> static const struct net_device_ops xenvif_netdev_ops = {
> + .ndo_select_queue = xenvif_select_queue,
> .ndo_start_xmit = xenvif_start_xmit,
> .ndo_get_stats = xenvif_get_stats,
> .ndo_open = xenvif_open,
> diff --git a/drivers/net/xen-netback/netback.c
> b/drivers/net/xen-netback/netback.c
> index a1f1a38..41ec7e9 100644
> --- a/drivers/net/xen-netback/netback.c
> +++ b/drivers/net/xen-netback/netback.c
> @@ -2163,6 +2163,89 @@ int xenvif_dealloc_kthread(void *data)
> return 0;
> }
>
> +static u32 xenvif_set_toeplitz_flags(struct xenvif *vif, u32 flags)
> +{
> + if (flags & ~(XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV4 |
> + XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV4_TCP |
> + XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV6 |
> + XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV6_TCP))
> + return XEN_NETIF_CTRL_STATUS_INVALID_PARAMETER;
> +
> + vif->toeplitz.flags = flags;
> +
> + return XEN_NETIF_CTRL_STATUS_SUCCESS;
> +}
> +
> +static u32 xenvif_set_toeplitz_key(struct xenvif *vif, u32 gref, u32 len)
> +{
> + u8 *key = vif->toeplitz.key;
> + struct gnttab_copy copy_op = {
> + .source.u.ref = gref,
> + .source.domid = vif->domid,
> + .dest.u.gmfn = virt_to_gfn(key),
> + .dest.domid = DOMID_SELF,
> + .dest.offset = xen_offset_in_page(key),
> + .len = len,
> + .flags = GNTCOPY_source_gref
> + };
> +
> + if (len > XEN_NETBK_MAX_TOEPLITZ_KEY_SIZE)
> + return XEN_NETIF_CTRL_STATUS_INVALID_PARAMETER;
> +
> + gnttab_batch_copy(©_op, 1);
> +
> + if (copy_op.status != GNTST_okay)
> + return XEN_NETIF_CTRL_STATUS_INVALID_PARAMETER;
> +
> + /* Clear any remaining key octets */
> + if (len < XEN_NETBK_MAX_TOEPLITZ_KEY_SIZE)
> + memset(key + len, 0, XEN_NETBK_MAX_TOEPLITZ_KEY_SIZE - len);
> +
> + return XEN_NETIF_CTRL_STATUS_SUCCESS;
> +}
> +
> +static u32 xenvif_set_toeplitz_mapping_order(struct xenvif *vif,
> + u32 order)
> +{
> + if (order > XEN_NETBK_MAX_TOEPLITZ_MAPPING_ORDER)
> + return XEN_NETIF_CTRL_STATUS_INVALID_PARAMETER;
> +
> + vif->toeplitz.order = order;
> + memset(vif->toeplitz.mapping, 0, sizeof(u32) << order);
> +
> + return XEN_NETIF_CTRL_STATUS_SUCCESS;
> +}
> +
> +static u32 xenvif_set_toeplitz_mapping(struct xenvif *vif, u32 gref,
> + u32 len, u32 off)
> +{
> + u32 *mapping = &vif->toeplitz.mapping[off];
> + struct gnttab_copy copy_op = {
> + .source.u.ref = gref,
> + .source.domid = vif->domid,
> + .dest.u.gmfn = virt_to_gfn(mapping),
> + .dest.domid = DOMID_SELF,
> + .dest.offset = xen_offset_in_page(mapping),
> + .len = len * sizeof(u32),
> + .flags = GNTCOPY_source_gref
> + };
> +
> + if ((off + len > (1u << vif->toeplitz.order)) ||
> + copy_op.len > XEN_PAGE_SIZE)
> + return XEN_NETIF_CTRL_STATUS_INVALID_PARAMETER;
> +
> + while (len-- != 0)
> + if (mapping[off++] >= vif->num_queues)
> + return XEN_NETIF_CTRL_STATUS_INVALID_PARAMETER;
> +
> + gnttab_batch_copy(©_op, 1);
> +
> + if (copy_op.status != GNTST_okay)
> + return XEN_NETIF_CTRL_STATUS_INVALID_PARAMETER;
> +
> + return XEN_NETIF_CTRL_STATUS_SUCCESS;
> +}
> +
> static void make_ctrl_response(struct xenvif *vif,
> const struct xen_netif_ctrl_request *req,
> u32 status, u32 data)
> @@ -2191,9 +2274,48 @@ static void push_ctrl_response(struct xenvif *vif)
> static void process_ctrl_request(struct xenvif *vif,
> const struct xen_netif_ctrl_request *req)
> {
> - /* There is no support for control requests yet. */
> - make_ctrl_response(vif, req,
> - XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED, 0);
> + u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
> + u32 data = 0;
> +
> + switch (req->type) {
> + case XEN_NETIF_CTRL_TYPE_GET_TOEPLITZ_FLAGS:
> + status = XEN_NETIF_CTRL_STATUS_SUCCESS;
> + data = XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV4 |
> + XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV4_TCP |
> + XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV6 |
> + XEN_NETIF_CTRL_TOEPLITZ_HASH_IPV6_TCP;
> + break;
> +
> + case XEN_NETIF_CTRL_TYPE_SET_TOEPLITZ_FLAGS:
> + status = xenvif_set_toeplitz_flags(vif, req->data[0]);
> + break;
> +
> + case XEN_NETIF_CTRL_TYPE_SET_TOEPLITZ_KEY:
> + status = xenvif_set_toeplitz_key(vif, req->data[0],
> + req->data[1]);
> + break;
> +
> + case XEN_NETIF_CTRL_TYPE_GET_TOEPLITZ_MAPPING_ORDER:
> + status = XEN_NETIF_CTRL_STATUS_SUCCESS;
> + data = XEN_NETBK_MAX_TOEPLITZ_MAPPING_ORDER;
> + break;
> +
> + case XEN_NETIF_CTRL_TYPE_SET_TOEPLITZ_MAPPING_ORDER:
> + status = xenvif_set_toeplitz_mapping_order(vif,
> + req->data[0]);
> + break;
> +
> + case XEN_NETIF_CTRL_TYPE_SET_TOEPLITZ_MAPPING:
> + status = xenvif_set_toeplitz_mapping(vif, req->data[0],
> + req->data[1],
> + req->data[2]);
> + break;
> +
> + default:
> + break;
> + }
> +
> + make_ctrl_response(vif, req, status, data);
> push_ctrl_response(vif);
> }
>
> --
> 2.1.4
>
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel
|
 |
Lists.xenproject.org is hosted with RackSpace, monitoring our |