netdevsim: add bpf offload support
authorJakub Kicinski <jakub.kicinski@netronome.com>
Fri, 1 Dec 2017 23:08:59 +0000 (15:08 -0800)
committerDaniel Borkmann <daniel@iogearbox.net>
Sat, 2 Dec 2017 23:27:57 +0000 (00:27 +0100)
Add support for loading programs for netdevsim devices and
expose the related information via DebugFS.  Both offload
of XDP and cls_bpf programs is supported.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
drivers/net/netdevsim/Makefile
drivers/net/netdevsim/bpf.c [new file with mode: 0644]
drivers/net/netdevsim/netdev.c
drivers/net/netdevsim/netdevsim.h

index 07867bfe873bc577cacfe78df2bc763c884bf8c7..074ddebbc41d3474cd90a2556e854140b9275aa5 100644 (file)
@@ -4,3 +4,4 @@ obj-$(CONFIG_NETDEVSIM) += netdevsim.o
 
 netdevsim-objs := \
        netdev.o \
+       bpf.o \
diff --git a/drivers/net/netdevsim/bpf.c b/drivers/net/netdevsim/bpf.c
new file mode 100644 (file)
index 0000000..8e4398a
--- /dev/null
@@ -0,0 +1,373 @@
+/*
+ * Copyright (C) 2017 Netronome Systems, Inc.
+ *
+ * This software is licensed under the GNU General License Version 2,
+ * June 1991 as shown in the file COPYING in the top-level directory of this
+ * source tree.
+ *
+ * THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS"
+ * WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING,
+ * BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE
+ * OF THE PROGRAM IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME
+ * THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+ */
+
+#include <linux/bpf.h>
+#include <linux/bpf_verifier.h>
+#include <linux/debugfs.h>
+#include <linux/kernel.h>
+#include <linux/rtnetlink.h>
+#include <net/pkt_cls.h>
+
+#include "netdevsim.h"
+
+struct nsim_bpf_bound_prog {
+       struct netdevsim *ns;
+       struct bpf_prog *prog;
+       struct dentry *ddir;
+       const char *state;
+       bool is_loaded;
+       struct list_head l;
+};
+
+static int nsim_debugfs_bpf_string_read(struct seq_file *file, void *data)
+{
+       const char **str = file->private;
+
+       if (*str)
+               seq_printf(file, "%s\n", *str);
+
+       return 0;
+}
+
+static int nsim_debugfs_bpf_string_open(struct inode *inode, struct file *f)
+{
+       return single_open(f, nsim_debugfs_bpf_string_read, inode->i_private);
+}
+
+static const struct file_operations nsim_bpf_string_fops = {
+       .owner = THIS_MODULE,
+       .open = nsim_debugfs_bpf_string_open,
+       .release = single_release,
+       .read = seq_read,
+       .llseek = seq_lseek
+};
+
+static int
+nsim_bpf_verify_insn(struct bpf_verifier_env *env, int insn_idx, int prev_insn)
+{
+       struct nsim_bpf_bound_prog *state;
+
+       state = env->prog->aux->offload->dev_priv;
+       if (state->ns->bpf_bind_verifier_delay && !insn_idx)
+               msleep(state->ns->bpf_bind_verifier_delay);
+
+       return 0;
+}
+
+static const struct bpf_ext_analyzer_ops nsim_bpf_analyzer_ops = {
+       .insn_hook = nsim_bpf_verify_insn,
+};
+
+static bool nsim_xdp_offload_active(struct netdevsim *ns)
+{
+       return ns->xdp_prog_mode == XDP_ATTACHED_HW;
+}
+
+static void nsim_prog_set_loaded(struct bpf_prog *prog, bool loaded)
+{
+       struct nsim_bpf_bound_prog *state;
+
+       if (!prog || !prog->aux->offload)
+               return;
+
+       state = prog->aux->offload->dev_priv;
+       state->is_loaded = loaded;
+}
+
+static int
+nsim_bpf_offload(struct netdevsim *ns, struct bpf_prog *prog, bool oldprog)
+{
+       nsim_prog_set_loaded(ns->bpf_offloaded, false);
+
+       WARN(!!ns->bpf_offloaded != oldprog,
+            "bad offload state, expected offload %sto be active",
+            oldprog ? "" : "not ");
+       ns->bpf_offloaded = prog;
+       ns->bpf_offloaded_id = prog ? prog->aux->id : 0;
+       nsim_prog_set_loaded(prog, true);
+
+       return 0;
+}
+
+int nsim_bpf_setup_tc_block_cb(enum tc_setup_type type,
+                              void *type_data, void *cb_priv)
+{
+       struct tc_cls_bpf_offload *cls_bpf = type_data;
+       struct bpf_prog *prog = cls_bpf->prog;
+       struct netdevsim *ns = cb_priv;
+       bool skip_sw;
+
+       if (type != TC_SETUP_CLSBPF ||
+           !tc_can_offload(ns->netdev) ||
+           cls_bpf->common.protocol != htons(ETH_P_ALL) ||
+           cls_bpf->common.chain_index)
+               return -EOPNOTSUPP;
+
+       skip_sw = cls_bpf->gen_flags & TCA_CLS_FLAGS_SKIP_SW;
+
+       if (nsim_xdp_offload_active(ns))
+               return -EBUSY;
+
+       if (!ns->bpf_tc_accept)
+               return -EOPNOTSUPP;
+       /* Note: progs without skip_sw will probably not be dev bound */
+       if (prog && !prog->aux->offload && !ns->bpf_tc_non_bound_accept)
+               return -EOPNOTSUPP;
+
+       switch (cls_bpf->command) {
+       case TC_CLSBPF_REPLACE:
+               return nsim_bpf_offload(ns, prog, true);
+       case TC_CLSBPF_ADD:
+               return nsim_bpf_offload(ns, prog, false);
+       case TC_CLSBPF_DESTROY:
+               return nsim_bpf_offload(ns, NULL, true);
+       default:
+               return -EOPNOTSUPP;
+       }
+}
+
+int nsim_bpf_disable_tc(struct netdevsim *ns)
+{
+       if (ns->bpf_offloaded && !nsim_xdp_offload_active(ns))
+               return -EBUSY;
+       return 0;
+}
+
+static int nsim_xdp_offload_prog(struct netdevsim *ns, struct netdev_bpf *bpf)
+{
+       if (!nsim_xdp_offload_active(ns) && !bpf->prog)
+               return 0;
+       if (!nsim_xdp_offload_active(ns) && bpf->prog && ns->bpf_offloaded) {
+               NSIM_EA(bpf->extack, "TC program is already loaded");
+               return -EBUSY;
+       }
+
+       return nsim_bpf_offload(ns, bpf->prog, nsim_xdp_offload_active(ns));
+}
+
+static int nsim_xdp_set_prog(struct netdevsim *ns, struct netdev_bpf *bpf)
+{
+       int err;
+
+       if (ns->xdp_prog && (bpf->flags ^ ns->xdp_flags) & XDP_FLAGS_MODES) {
+               NSIM_EA(bpf->extack, "program loaded with different flags");
+               return -EBUSY;
+       }
+
+       if (bpf->command == XDP_SETUP_PROG && !ns->bpf_xdpdrv_accept) {
+               NSIM_EA(bpf->extack, "driver XDP disabled in DebugFS");
+               return -EOPNOTSUPP;
+       }
+       if (bpf->command == XDP_SETUP_PROG_HW && !ns->bpf_xdpoffload_accept) {
+               NSIM_EA(bpf->extack, "XDP offload disabled in DebugFS");
+               return -EOPNOTSUPP;
+       }
+
+       if (bpf->command == XDP_SETUP_PROG_HW) {
+               err = nsim_xdp_offload_prog(ns, bpf);
+               if (err)
+                       return err;
+       }
+
+       if (ns->xdp_prog)
+               bpf_prog_put(ns->xdp_prog);
+
+       ns->xdp_prog = bpf->prog;
+       ns->xdp_flags = bpf->flags;
+
+       if (!bpf->prog)
+               ns->xdp_prog_mode = XDP_ATTACHED_NONE;
+       else if (bpf->command == XDP_SETUP_PROG)
+               ns->xdp_prog_mode = XDP_ATTACHED_DRV;
+       else
+               ns->xdp_prog_mode = XDP_ATTACHED_HW;
+
+       return 0;
+}
+
+int nsim_bpf_create_prog(struct netdevsim *ns, struct bpf_prog *prog)
+{
+       struct nsim_bpf_bound_prog *state;
+       char name[16];
+       int err;
+
+       state = kzalloc(sizeof(*state), GFP_KERNEL);
+       if (!state)
+               return -ENOMEM;
+
+       state->ns = ns;
+       state->prog = prog;
+       state->state = "verify";
+
+       /* Program id is not populated yet when we create the state. */
+       sprintf(name, "%u", ns->prog_id_gen++);
+       state->ddir = debugfs_create_dir(name, ns->ddir_bpf_bound_progs);
+       if (IS_ERR(state->ddir)) {
+               err = PTR_ERR(state->ddir);
+               kfree(state);
+               return err;
+       }
+
+       debugfs_create_u32("id", 0400, state->ddir, &prog->aux->id);
+       debugfs_create_file("state", 0400, state->ddir,
+                           &state->state, &nsim_bpf_string_fops);
+       debugfs_create_bool("loaded", 0400, state->ddir, &state->is_loaded);
+
+       list_add_tail(&state->l, &ns->bpf_bound_progs);
+
+       prog->aux->offload->dev_priv = state;
+
+       return 0;
+}
+
+void nsim_bpf_destroy_prog(struct bpf_prog *prog)
+{
+       struct nsim_bpf_bound_prog *state;
+
+       state = prog->aux->offload->dev_priv;
+       WARN(state->is_loaded,
+            "offload state destroyed while program still bound");
+       debugfs_remove_recursive(state->ddir);
+       list_del(&state->l);
+       kfree(state);
+}
+
+static int nsim_setup_prog_checks(struct netdevsim *ns, struct netdev_bpf *bpf)
+{
+       if (bpf->prog && bpf->prog->aux->offload) {
+               NSIM_EA(bpf->extack, "attempt to load offloaded prog to drv");
+               return -EINVAL;
+       }
+       if (ns->netdev->mtu > NSIM_XDP_MAX_MTU) {
+               NSIM_EA(bpf->extack, "MTU too large w/ XDP enabled");
+               return -EINVAL;
+       }
+       if (nsim_xdp_offload_active(ns)) {
+               NSIM_EA(bpf->extack, "xdp offload active, can't load drv prog");
+               return -EBUSY;
+       }
+       return 0;
+}
+
+static int
+nsim_setup_prog_hw_checks(struct netdevsim *ns, struct netdev_bpf *bpf)
+{
+       struct nsim_bpf_bound_prog *state;
+
+       if (!bpf->prog)
+               return 0;
+
+       if (!bpf->prog->aux->offload) {
+               NSIM_EA(bpf->extack, "xdpoffload of non-bound program");
+               return -EINVAL;
+       }
+       if (bpf->prog->aux->offload->netdev != ns->netdev) {
+               NSIM_EA(bpf->extack, "program bound to different dev");
+               return -EINVAL;
+       }
+
+       state = bpf->prog->aux->offload->dev_priv;
+       if (WARN_ON(strcmp(state->state, "xlated"))) {
+               NSIM_EA(bpf->extack, "offloading program in bad state");
+               return -EINVAL;
+       }
+       return 0;
+}
+
+int nsim_bpf(struct net_device *dev, struct netdev_bpf *bpf)
+{
+       struct netdevsim *ns = netdev_priv(dev);
+       struct nsim_bpf_bound_prog *state;
+       int err;
+
+       ASSERT_RTNL();
+
+       switch (bpf->command) {
+       case BPF_OFFLOAD_VERIFIER_PREP:
+               if (!ns->bpf_bind_accept)
+                       return -EOPNOTSUPP;
+
+               err = nsim_bpf_create_prog(ns, bpf->verifier.prog);
+               if (err)
+                       return err;
+
+               bpf->verifier.ops = &nsim_bpf_analyzer_ops;
+               return 0;
+       case BPF_OFFLOAD_TRANSLATE:
+               state = bpf->offload.prog->aux->offload->dev_priv;
+
+               state->state = "xlated";
+               return 0;
+       case BPF_OFFLOAD_DESTROY:
+               nsim_bpf_destroy_prog(bpf->offload.prog);
+               return 0;
+       case XDP_QUERY_PROG:
+               bpf->prog_attached = ns->xdp_prog_mode;
+               bpf->prog_id = ns->xdp_prog ? ns->xdp_prog->aux->id : 0;
+               bpf->prog_flags = ns->xdp_prog ? ns->xdp_flags : 0;
+               return 0;
+       case XDP_SETUP_PROG:
+               err = nsim_setup_prog_checks(ns, bpf);
+               if (err)
+                       return err;
+
+               return nsim_xdp_set_prog(ns, bpf);
+       case XDP_SETUP_PROG_HW:
+               err = nsim_setup_prog_hw_checks(ns, bpf);
+               if (err)
+                       return err;
+
+               return nsim_xdp_set_prog(ns, bpf);
+       default:
+               return -EINVAL;
+       }
+}
+
+int nsim_bpf_init(struct netdevsim *ns)
+{
+       INIT_LIST_HEAD(&ns->bpf_bound_progs);
+
+       debugfs_create_u32("bpf_offloaded_id", 0400, ns->ddir,
+                          &ns->bpf_offloaded_id);
+
+       ns->bpf_bind_accept = true;
+       debugfs_create_bool("bpf_bind_accept", 0600, ns->ddir,
+                           &ns->bpf_bind_accept);
+       debugfs_create_u32("bpf_bind_verifier_delay", 0600, ns->ddir,
+                          &ns->bpf_bind_verifier_delay);
+       ns->ddir_bpf_bound_progs =
+               debugfs_create_dir("bpf_bound_progs", ns->ddir);
+
+       ns->bpf_tc_accept = true;
+       debugfs_create_bool("bpf_tc_accept", 0600, ns->ddir,
+                           &ns->bpf_tc_accept);
+       debugfs_create_bool("bpf_tc_non_bound_accept", 0600, ns->ddir,
+                           &ns->bpf_tc_non_bound_accept);
+       ns->bpf_xdpdrv_accept = true;
+       debugfs_create_bool("bpf_xdpdrv_accept", 0600, ns->ddir,
+                           &ns->bpf_xdpdrv_accept);
+       ns->bpf_xdpoffload_accept = true;
+       debugfs_create_bool("bpf_xdpoffload_accept", 0600, ns->ddir,
+                           &ns->bpf_xdpoffload_accept);
+
+       return 0;
+}
+
+void nsim_bpf_uninit(struct netdevsim *ns)
+{
+       WARN_ON(!list_empty(&ns->bpf_bound_progs));
+       WARN_ON(ns->xdp_prog);
+       WARN_ON(ns->bpf_offloaded);
+}
index 7599c72c477a3df25009acec3d7e829b304acd3b..828c1ce49a8b4a429b10714b6d9692e0bc816b90 100644 (file)
  * THE COST OF ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
  */
 
+#include <linux/debugfs.h>
 #include <linux/etherdevice.h>
 #include <linux/kernel.h>
 #include <linux/module.h>
 #include <linux/netdevice.h>
 #include <linux/slab.h>
 #include <net/netlink.h>
+#include <net/pkt_cls.h>
 #include <net/rtnetlink.h>
 
 #include "netdevsim.h"
 
+static int nsim_init(struct net_device *dev)
+{
+       struct netdevsim *ns = netdev_priv(dev);
+       int err;
+
+       ns->netdev = dev;
+       ns->ddir = debugfs_create_dir(netdev_name(dev), nsim_ddir);
+
+       err = nsim_bpf_init(ns);
+       if (err)
+               goto err_debugfs_destroy;
+
+       return 0;
+
+err_debugfs_destroy:
+       debugfs_remove_recursive(ns->ddir);
+       return err;
+}
+
+static void nsim_uninit(struct net_device *dev)
+{
+       struct netdevsim *ns = netdev_priv(dev);
+
+       debugfs_remove_recursive(ns->ddir);
+       nsim_bpf_uninit(ns);
+}
+
 static netdev_tx_t nsim_start_xmit(struct sk_buff *skb, struct net_device *dev)
 {
        struct netdevsim *ns = netdev_priv(dev);
@@ -41,6 +70,19 @@ static void nsim_set_rx_mode(struct net_device *dev)
 {
 }
 
+static int nsim_change_mtu(struct net_device *dev, int new_mtu)
+{
+       struct netdevsim *ns = netdev_priv(dev);
+
+       if (ns->xdp_prog_mode == XDP_ATTACHED_DRV &&
+           new_mtu > NSIM_XDP_MAX_MTU)
+               return -EBUSY;
+
+       dev->mtu = new_mtu;
+
+       return 0;
+}
+
 static void
 nsim_get_stats64(struct net_device *dev, struct rtnl_link_stats64 *stats)
 {
@@ -54,12 +96,66 @@ nsim_get_stats64(struct net_device *dev, struct rtnl_link_stats64 *stats)
        } while (u64_stats_fetch_retry(&ns->syncp, start));
 }
 
+static int
+nsim_setup_tc_block_cb(enum tc_setup_type type, void *type_data, void *cb_priv)
+{
+       return nsim_bpf_setup_tc_block_cb(type, type_data, cb_priv);
+}
+
+static int
+nsim_setup_tc_block(struct net_device *dev, struct tc_block_offload *f)
+{
+       struct netdevsim *ns = netdev_priv(dev);
+
+       if (f->binder_type != TCF_BLOCK_BINDER_TYPE_CLSACT_INGRESS)
+               return -EOPNOTSUPP;
+
+       switch (f->command) {
+       case TC_BLOCK_BIND:
+               return tcf_block_cb_register(f->block, nsim_setup_tc_block_cb,
+                                            ns, ns);
+       case TC_BLOCK_UNBIND:
+               tcf_block_cb_unregister(f->block, nsim_setup_tc_block_cb, ns);
+               return 0;
+       default:
+               return -EOPNOTSUPP;
+       }
+}
+
+static int
+nsim_setup_tc(struct net_device *dev, enum tc_setup_type type, void *type_data)
+{
+       switch (type) {
+       case TC_SETUP_BLOCK:
+               return nsim_setup_tc_block(dev, type_data);
+       default:
+               return -EOPNOTSUPP;
+       }
+}
+
+static int
+nsim_set_features(struct net_device *dev, netdev_features_t features)
+{
+       struct netdevsim *ns = netdev_priv(dev);
+
+       if ((dev->features & NETIF_F_HW_TC) > (features & NETIF_F_HW_TC))
+               return nsim_bpf_disable_tc(ns);
+
+       return 0;
+}
+
 static const struct net_device_ops nsim_netdev_ops = {
+       .ndo_init               = nsim_init,
+       .ndo_uninit             = nsim_uninit,
        .ndo_start_xmit         = nsim_start_xmit,
        .ndo_set_rx_mode        = nsim_set_rx_mode,
        .ndo_set_mac_address    = eth_mac_addr,
        .ndo_validate_addr      = eth_validate_addr,
+       .ndo_change_mtu         = nsim_change_mtu,
        .ndo_get_stats64        = nsim_get_stats64,
+       .ndo_setup_tc           = nsim_setup_tc,
+       .ndo_set_features       = nsim_set_features,
+       .ndo_bpf                = nsim_bpf,
 };
 
 static void nsim_setup(struct net_device *dev)
@@ -80,6 +176,7 @@ static void nsim_setup(struct net_device *dev)
                         NETIF_F_FRAGLIST |
                         NETIF_F_HW_CSUM |
                         NETIF_F_TSO;
+       dev->hw_features |= NETIF_F_HW_TC;
        dev->max_mtu = ETH_MAX_MTU;
 }
 
@@ -102,14 +199,31 @@ static struct rtnl_link_ops nsim_link_ops __read_mostly = {
        .validate       = nsim_validate,
 };
 
+struct dentry *nsim_ddir;
+
 static int __init nsim_module_init(void)
 {
-       return rtnl_link_register(&nsim_link_ops);
+       int err;
+
+       nsim_ddir = debugfs_create_dir(DRV_NAME, NULL);
+       if (IS_ERR(nsim_ddir))
+               return PTR_ERR(nsim_ddir);
+
+       err = rtnl_link_register(&nsim_link_ops);
+       if (err)
+               goto err_debugfs_destroy;
+
+       return 0;
+
+err_debugfs_destroy:
+       debugfs_remove_recursive(nsim_ddir);
+       return err;
 }
 
 static void __exit nsim_module_exit(void)
 {
        rtnl_link_unregister(&nsim_link_ops);
+       debugfs_remove_recursive(nsim_ddir);
 }
 
 module_init(nsim_module_init);
index 4558c6f11598d789ebb8aeb517b8611adda5ea8f..8779e6a8f885cc43af0407b8d8356f071f3e21f4 100644 (file)
  */
 
 #include <linux/kernel.h>
+#include <linux/list.h>
 #include <linux/netdevice.h>
 #include <linux/u64_stats_sync.h>
 
 #define DRV_NAME       "netdevsim"
 
+#define NSIM_XDP_MAX_MTU       4000
+
+#define NSIM_EA(extack, msg)   NL_SET_ERR_MSG_MOD((extack), msg)
+
+struct bpf_prog;
+struct dentry;
+
 struct netdevsim {
+       struct net_device *netdev;
+
        u64 tx_packets;
        u64 tx_bytes;
        struct u64_stats_sync syncp;
+
+       struct dentry *ddir;
+
+       struct bpf_prog *bpf_offloaded;
+       u32 bpf_offloaded_id;
+
+       u32 xdp_flags;
+       int xdp_prog_mode;
+       struct bpf_prog *xdp_prog;
+
+       u32 prog_id_gen;
+
+       bool bpf_bind_accept;
+       u32 bpf_bind_verifier_delay;
+       struct dentry *ddir_bpf_bound_progs;
+       struct list_head bpf_bound_progs;
+
+       bool bpf_tc_accept;
+       bool bpf_tc_non_bound_accept;
+       bool bpf_xdpdrv_accept;
+       bool bpf_xdpoffload_accept;
 };
+
+extern struct dentry *nsim_ddir;
+
+int nsim_bpf_init(struct netdevsim *ns);
+void nsim_bpf_uninit(struct netdevsim *ns);
+int nsim_bpf(struct net_device *dev, struct netdev_bpf *bpf);
+int nsim_bpf_disable_tc(struct netdevsim *ns);
+int nsim_bpf_setup_tc_block_cb(enum tc_setup_type type,
+                              void *type_data, void *cb_priv);