net: add network priority cgroup infrastructure (v4)
authorNeil Horman <nhorman@tuxdriver.com>
Tue, 22 Nov 2011 05:10:51 +0000 (05:10 +0000)
committerDavid S. Miller <davem@davemloft.net>
Tue, 22 Nov 2011 20:22:23 +0000 (15:22 -0500)
This patch adds in the infrastructure code to create the network priority
cgroup.  The cgroup, in addition to the standard processes file creates two
control files:

1) prioidx - This is a read-only file that exports the index of this cgroup.
This is a value that is both arbitrary and unique to a cgroup in this subsystem,
and is used to index the per-device priority map

2) priomap - This is a writeable file.  On read it reports a table of 2-tuples
<name:priority> where name is the name of a network interface and priority is
indicates the priority assigned to frames egresessing on the named interface and
originating from a pid in this cgroup

This cgroup allows for skb priority to be set prior to a root qdisc getting
selected. This is benenficial for DCB enabled systems, in that it allows for any
application to use dcb configured priorities so without application modification

Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
CC: Robert Love <robert.w.love@intel.com>
CC: "David S. Miller" <davem@davemloft.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/cgroup_subsys.h
include/linux/netdevice.h
include/net/netprio_cgroup.h [new file with mode: 0644]
include/net/sock.h
net/Kconfig
net/core/Makefile
net/core/dev.c
net/core/netprio_cgroup.c [new file with mode: 0644]
net/core/sock.c
net/socket.c

index ac663c1..0bd390c 100644 (file)
@@ -59,8 +59,16 @@ SUBSYS(net_cls)
 SUBSYS(blkio)
 #endif
 
+/* */
+
 #ifdef CONFIG_CGROUP_PERF
 SUBSYS(perf)
 #endif
 
 /* */
+
+#ifdef CONFIG_NETPRIO_CGROUP
+SUBSYS(net_prio)
+#endif
+
+/* */
index 3eb383a..999bb26 100644 (file)
@@ -50,6 +50,7 @@
 #ifdef CONFIG_DCB
 #include <net/dcbnl.h>
 #endif
+#include <net/netprio_cgroup.h>
 
 #include <linux/netdev_features.h>
 
@@ -1245,6 +1246,9 @@ struct net_device {
        /* max exchange id for FCoE LRO by ddp */
        unsigned int            fcoe_ddp_xid;
 #endif
+#if IS_ENABLED(CONFIG_NETPRIO_CGROUP)
+       struct netprio_map __rcu *priomap;
+#endif
        /* phy device may attach itself for hardware timestamping */
        struct phy_device *phydev;
 
diff --git a/include/net/netprio_cgroup.h b/include/net/netprio_cgroup.h
new file mode 100644 (file)
index 0000000..c432e99
--- /dev/null
@@ -0,0 +1,65 @@
+/*
+ * netprio_cgroup.h                    Control Group Priority set
+ *
+ *
+ * Authors:    Neil Horman <nhorman@tuxdriver.com>
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License as published by the Free
+ * Software Foundation; either version 2 of the License, or (at your option)
+ * any later version.
+ *
+ */
+
+#ifndef _NETPRIO_CGROUP_H
+#define _NETPRIO_CGROUP_H
+#include <linux/module.h>
+#include <linux/cgroup.h>
+#include <linux/hardirq.h>
+#include <linux/rcupdate.h>
+
+struct cgroup_netprio_state
+{
+       struct cgroup_subsys_state css;
+       u32 prioidx;
+};
+
+struct netprio_map {
+       struct rcu_head rcu;
+       u32 priomap_len;
+       u32 priomap[];
+};
+
+#ifdef CONFIG_CGROUPS
+
+#ifndef CONFIG_NETPRIO_CGROUP
+extern int net_prio_subsys_id;
+#endif
+
+extern void sock_update_netprioidx(struct sock *sk);
+
+static inline struct cgroup_netprio_state
+               *task_netprio_state(struct task_struct *p)
+{
+#if IS_ENABLED(CONFIG_NETPRIO_CGROUP)
+       return container_of(task_subsys_state(p, net_prio_subsys_id),
+                           struct cgroup_netprio_state, css);
+#else
+       return NULL;
+#endif
+}
+
+#else
+
+#define sock_update_netprioidx(sk)
+#define skb_update_prio(skb)
+
+static inline struct cgroup_netprio_state
+               *task_netprio_state(struct task_struct *p)
+{
+       return NULL;
+}
+
+#endif
+
+#endif  /* _NET_CLS_CGROUP_H */
index 1c28f39..8ac338c 100644 (file)
@@ -320,6 +320,9 @@ struct sock {
        unsigned short          sk_ack_backlog;
        unsigned short          sk_max_ack_backlog;
        __u32                   sk_priority;
+#ifdef CONFIG_CGROUPS
+       __u32                   sk_cgrp_prioidx;
+#endif
        struct pid              *sk_peer_pid;
        const struct cred       *sk_peer_cred;
        long                    sk_rcvtimeo;
index a073148..63d2c5d 100644 (file)
@@ -232,6 +232,13 @@ config XPS
        depends on SMP && SYSFS && USE_GENERIC_SMP_HELPERS
        default y
 
+config NETPRIO_CGROUP
+       tristate "Network priority cgroup"
+       depends on CGROUPS
+       ---help---
+         Cgroup subsystem for use in assigning processes to network priorities on
+         a per-interface basis
+
 config HAVE_BPF_JIT
        bool
 
index 0d357b1..3606d40 100644 (file)
@@ -19,3 +19,4 @@ obj-$(CONFIG_FIB_RULES) += fib_rules.o
 obj-$(CONFIG_TRACEPOINTS) += net-traces.o
 obj-$(CONFIG_NET_DROP_MONITOR) += drop_monitor.o
 obj-$(CONFIG_NETWORK_PHY_TIMESTAMPING) += timestamping.o
+obj-$(CONFIG_NETPRIO_CGROUP) += netprio_cgroup.o
index f789599..8afb244 100644 (file)
@@ -2449,6 +2449,18 @@ static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
        return rc;
 }
 
+#if IS_ENABLED(CONFIG_NETPRIO_CGROUP)
+static void skb_update_prio(struct sk_buff *skb)
+{
+       struct netprio_map *map = rcu_dereference(skb->dev->priomap);
+
+       if ((!skb->priority) && (skb->sk) && map)
+               skb->priority = map->priomap[skb->sk->sk_cgrp_prioidx];
+}
+#else
+#define skb_update_prio(skb)
+#endif
+
 static DEFINE_PER_CPU(int, xmit_recursion);
 #define RECURSION_LIMIT 10
 
@@ -2489,6 +2501,8 @@ int dev_queue_xmit(struct sk_buff *skb)
         */
        rcu_read_lock_bh();
 
+       skb_update_prio(skb);
+
        txq = dev_pick_tx(dev, skb);
        q = rcu_dereference_bh(txq->qdisc);
 
diff --git a/net/core/netprio_cgroup.c b/net/core/netprio_cgroup.c
new file mode 100644 (file)
index 0000000..72ad0bc
--- /dev/null
@@ -0,0 +1,344 @@
+/*
+ * net/core/netprio_cgroup.c   Priority Control Group
+ *
+ *             This program is free software; you can redistribute it and/or
+ *             modify it under the terms of the GNU General Public License
+ *             as published by the Free Software Foundation; either version
+ *             2 of the License, or (at your option) any later version.
+ *
+ * Authors:    Neil Horman <nhorman@tuxdriver.com>
+ */
+
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <linux/types.h>
+#include <linux/string.h>
+#include <linux/errno.h>
+#include <linux/skbuff.h>
+#include <linux/cgroup.h>
+#include <linux/rcupdate.h>
+#include <linux/atomic.h>
+#include <net/rtnetlink.h>
+#include <net/pkt_cls.h>
+#include <net/sock.h>
+#include <net/netprio_cgroup.h>
+
+static struct cgroup_subsys_state *cgrp_create(struct cgroup_subsys *ss,
+                                              struct cgroup *cgrp);
+static void cgrp_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp);
+static int cgrp_populate(struct cgroup_subsys *ss, struct cgroup *cgrp);
+
+struct cgroup_subsys net_prio_subsys = {
+       .name           = "net_prio",
+       .create         = cgrp_create,
+       .destroy        = cgrp_destroy,
+       .populate       = cgrp_populate,
+#ifdef CONFIG_NETPRIO_CGROUP
+       .subsys_id      = net_prio_subsys_id,
+#endif
+       .module         = THIS_MODULE
+};
+
+#define PRIOIDX_SZ 128
+
+static unsigned long prioidx_map[PRIOIDX_SZ];
+static DEFINE_SPINLOCK(prioidx_map_lock);
+static atomic_t max_prioidx = ATOMIC_INIT(0);
+
+static inline struct cgroup_netprio_state *cgrp_netprio_state(struct cgroup *cgrp)
+{
+       return container_of(cgroup_subsys_state(cgrp, net_prio_subsys_id),
+                           struct cgroup_netprio_state, css);
+}
+
+static int get_prioidx(u32 *prio)
+{
+       unsigned long flags;
+       u32 prioidx;
+
+       spin_lock_irqsave(&prioidx_map_lock, flags);
+       prioidx = find_first_zero_bit(prioidx_map, sizeof(unsigned long) * PRIOIDX_SZ);
+       set_bit(prioidx, prioidx_map);
+       spin_unlock_irqrestore(&prioidx_map_lock, flags);
+       if (prioidx == sizeof(unsigned long) * PRIOIDX_SZ)
+               return -ENOSPC;
+
+       atomic_set(&max_prioidx, prioidx);
+       *prio = prioidx;
+       return 0;
+}
+
+static void put_prioidx(u32 idx)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave(&prioidx_map_lock, flags);
+       clear_bit(idx, prioidx_map);
+       spin_unlock_irqrestore(&prioidx_map_lock, flags);
+}
+
+static void extend_netdev_table(struct net_device *dev, u32 new_len)
+{
+       size_t new_size = sizeof(struct netprio_map) +
+                          ((sizeof(u32) * new_len));
+       struct netprio_map *new_priomap = kzalloc(new_size, GFP_KERNEL);
+       struct netprio_map *old_priomap;
+       int i;
+
+       old_priomap  = rtnl_dereference(dev->priomap);
+
+       if (!new_priomap) {
+               printk(KERN_WARNING "Unable to alloc new priomap!\n");
+               return;
+       }
+
+       for (i = 0;
+            old_priomap && (i < old_priomap->priomap_len);
+            i++)
+               new_priomap->priomap[i] = old_priomap->priomap[i];
+
+       new_priomap->priomap_len = new_len;
+
+       rcu_assign_pointer(dev->priomap, new_priomap);
+       if (old_priomap)
+               kfree_rcu(old_priomap, rcu);
+}
+
+static void update_netdev_tables(void)
+{
+       struct net_device *dev;
+       u32 max_len = atomic_read(&max_prioidx);
+       struct netprio_map *map;
+
+       rtnl_lock();
+       for_each_netdev(&init_net, dev) {
+               map = rtnl_dereference(dev->priomap);
+               if ((!map) ||
+                   (map->priomap_len < max_len))
+                       extend_netdev_table(dev, max_len);
+       }
+       rtnl_unlock();
+}
+
+static struct cgroup_subsys_state *cgrp_create(struct cgroup_subsys *ss,
+                                                struct cgroup *cgrp)
+{
+       struct cgroup_netprio_state *cs;
+       int ret;
+
+       cs = kzalloc(sizeof(*cs), GFP_KERNEL);
+       if (!cs)
+               return ERR_PTR(-ENOMEM);
+
+       if (cgrp->parent && cgrp_netprio_state(cgrp->parent)->prioidx) {
+               kfree(cs);
+               return ERR_PTR(-EINVAL);
+       }
+
+       ret = get_prioidx(&cs->prioidx);
+       if (ret != 0) {
+               printk(KERN_WARNING "No space in priority index array\n");
+               kfree(cs);
+               return ERR_PTR(ret);
+       }
+
+       return &cs->css;
+}
+
+static void cgrp_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
+{
+       struct cgroup_netprio_state *cs;
+       struct net_device *dev;
+       struct netprio_map *map;
+
+       cs = cgrp_netprio_state(cgrp);
+       rtnl_lock();
+       for_each_netdev(&init_net, dev) {
+               map = rtnl_dereference(dev->priomap);
+               if (map)
+                       map->priomap[cs->prioidx] = 0;
+       }
+       rtnl_unlock();
+       put_prioidx(cs->prioidx);
+       kfree(cs);
+}
+
+static u64 read_prioidx(struct cgroup *cgrp, struct cftype *cft)
+{
+       return (u64)cgrp_netprio_state(cgrp)->prioidx;
+}
+
+static int read_priomap(struct cgroup *cont, struct cftype *cft,
+                       struct cgroup_map_cb *cb)
+{
+       struct net_device *dev;
+       u32 prioidx = cgrp_netprio_state(cont)->prioidx;
+       u32 priority;
+       struct netprio_map *map;
+
+       rcu_read_lock();
+       for_each_netdev_rcu(&init_net, dev) {
+               map = rcu_dereference(dev->priomap);
+               priority = map ? map->priomap[prioidx] : 0;
+               cb->fill(cb, dev->name, priority);
+       }
+       rcu_read_unlock();
+       return 0;
+}
+
+static int write_priomap(struct cgroup *cgrp, struct cftype *cft,
+                        const char *buffer)
+{
+       char *devname = kstrdup(buffer, GFP_KERNEL);
+       int ret = -EINVAL;
+       u32 prioidx = cgrp_netprio_state(cgrp)->prioidx;
+       unsigned long priority;
+       char *priostr;
+       struct net_device *dev;
+       struct netprio_map *map;
+
+       if (!devname)
+               return -ENOMEM;
+
+       /*
+        * Minimally sized valid priomap string
+        */
+       if (strlen(devname) < 3)
+               goto out_free_devname;
+
+       priostr = strstr(devname, " ");
+       if (!priostr)
+               goto out_free_devname;
+
+       /*
+        *Separate the devname from the associated priority
+        *and advance the priostr poitner to the priority value
+        */
+       *priostr = '\0';
+       priostr++;
+
+       /*
+        * If the priostr points to NULL, we're at the end of the passed
+        * in string, and its not a valid write
+        */
+       if (*priostr == '\0')
+               goto out_free_devname;
+
+       ret = kstrtoul(priostr, 10, &priority);
+       if (ret < 0)
+               goto out_free_devname;
+
+       ret = -ENODEV;
+
+       dev = dev_get_by_name(&init_net, devname);
+       if (!dev)
+               goto out_free_devname;
+
+       update_netdev_tables();
+       ret = 0;
+       rcu_read_lock();
+       map = rcu_dereference(dev->priomap);
+       if (map)
+               map->priomap[prioidx] = priority;
+       rcu_read_unlock();
+       dev_put(dev);
+
+out_free_devname:
+       kfree(devname);
+       return ret;
+}
+
+static struct cftype ss_files[] = {
+       {
+               .name = "prioidx",
+               .read_u64 = read_prioidx,
+       },
+       {
+               .name = "ifpriomap",
+               .read_map = read_priomap,
+               .write_string = write_priomap,
+       },
+};
+
+static int cgrp_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
+{
+       return cgroup_add_files(cgrp, ss, ss_files, ARRAY_SIZE(ss_files));
+}
+
+static int netprio_device_event(struct notifier_block *unused,
+                               unsigned long event, void *ptr)
+{
+       struct net_device *dev = ptr;
+       struct netprio_map *old;
+       u32 max_len = atomic_read(&max_prioidx);
+
+       /*
+        * Note this is called with rtnl_lock held so we have update side
+        * protection on our rcu assignments
+        */
+
+       switch (event) {
+
+       case NETDEV_REGISTER:
+               if (max_len)
+                       extend_netdev_table(dev, max_len);
+               break;
+       case NETDEV_UNREGISTER:
+               old = rtnl_dereference(dev->priomap);
+               rcu_assign_pointer(dev->priomap, NULL);
+               if (old)
+                       kfree_rcu(old, rcu);
+               break;
+       }
+       return NOTIFY_DONE;
+}
+
+static struct notifier_block netprio_device_notifier = {
+       .notifier_call = netprio_device_event
+};
+
+static int __init init_cgroup_netprio(void)
+{
+       int ret;
+
+       ret = cgroup_load_subsys(&net_prio_subsys);
+       if (ret)
+               goto out;
+#ifndef CONFIG_NETPRIO_CGROUP
+       smp_wmb();
+       net_prio_subsys_id = net_prio_subsys.subsys_id;
+#endif
+
+       register_netdevice_notifier(&netprio_device_notifier);
+
+out:
+       return ret;
+}
+
+static void __exit exit_cgroup_netprio(void)
+{
+       struct netprio_map *old;
+       struct net_device *dev;
+
+       unregister_netdevice_notifier(&netprio_device_notifier);
+
+       cgroup_unload_subsys(&net_prio_subsys);
+
+#ifndef CONFIG_NETPRIO_CGROUP
+       net_prio_subsys_id = -1;
+       synchronize_rcu();
+#endif
+
+       rtnl_lock();
+       for_each_netdev(&init_net, dev) {
+               old = rtnl_dereference(dev->priomap);
+               rcu_assign_pointer(dev->priomap, NULL);
+               if (old)
+                       kfree_rcu(old, rcu);
+       }
+       rtnl_unlock();
+}
+
+module_init(init_cgroup_netprio);
+module_exit(exit_cgroup_netprio);
+MODULE_LICENSE("GPL v2");
index 9a8b3fa..1606913 100644 (file)
 #include <net/xfrm.h>
 #include <linux/ipsec.h>
 #include <net/cls_cgroup.h>
+#include <net/netprio_cgroup.h>
 
 #include <linux/filter.h>
 
@@ -221,10 +222,16 @@ __u32 sysctl_rmem_default __read_mostly = SK_RMEM_MAX;
 int sysctl_optmem_max __read_mostly = sizeof(unsigned long)*(2*UIO_MAXIOV+512);
 EXPORT_SYMBOL(sysctl_optmem_max);
 
-#if defined(CONFIG_CGROUPS) && !defined(CONFIG_NET_CLS_CGROUP)
+#if defined(CONFIG_CGROUPS)
+#if !defined(CONFIG_NET_CLS_CGROUP)
 int net_cls_subsys_id = -1;
 EXPORT_SYMBOL_GPL(net_cls_subsys_id);
 #endif
+#if !defined(CONFIG_NETPRIO_CGROUP)
+int net_prio_subsys_id = -1;
+EXPORT_SYMBOL_GPL(net_prio_subsys_id);
+#endif
+#endif
 
 static int sock_set_timeout(long *timeo_p, char __user *optval, int optlen)
 {
@@ -1120,6 +1127,18 @@ void sock_update_classid(struct sock *sk)
                sk->sk_classid = classid;
 }
 EXPORT_SYMBOL(sock_update_classid);
+
+void sock_update_netprioidx(struct sock *sk)
+{
+       struct cgroup_netprio_state *state;
+       if (in_interrupt())
+               return;
+       rcu_read_lock();
+       state = task_netprio_state(current);
+       sk->sk_cgrp_prioidx = state ? state->prioidx : 0;
+       rcu_read_unlock();
+}
+EXPORT_SYMBOL_GPL(sock_update_netprioidx);
 #endif
 
 /**
@@ -1147,6 +1166,7 @@ struct sock *sk_alloc(struct net *net, int family, gfp_t priority,
                atomic_set(&sk->sk_wmem_alloc, 1);
 
                sock_update_classid(sk);
+               sock_update_netprioidx(sk);
        }
 
        return sk;
index 425ef42..e62b4f0 100644 (file)
@@ -551,6 +551,8 @@ static inline int __sock_sendmsg_nosec(struct kiocb *iocb, struct socket *sock,
 
        sock_update_classid(sock->sk);
 
+       sock_update_netprioidx(sock->sk);
+
        si->sock = sock;
        si->scm = NULL;
        si->msg = msg;