Minki/linux
1
0
Fork 1
mirror of https://github.com/torvalds/linux.git synced 2024-11-16 09:02:00 +00:00
Code Issues Packages Projects Releases Wiki Activity
5b8d8fb0fc
linux/drivers/net/eql.c
Eric Dumazet 0287587884 net: better IFF_XMIT_DST_RELEASE support 
Testing xmit_more support with netperf and connected UDP sockets,
I found strange dst refcount false sharing.

Current handling of IFF_XMIT_DST_RELEASE is not optimal.

Dropping dst in validate_xmit_skb() is certainly too late in case
packet was queued by cpu X but dequeued by cpu Y

The logical point to take care of drop/force is in __dev_queue_xmit()
before even taking qdisc lock.

As Julian Anastasov pointed out, need for skb_dst() might come from some
packet schedulers or classifiers.

This patch adds new helper to cleanly express needs of various drivers
or qdiscs/classifiers.

Drivers that need skb_dst() in their ndo_start_xmit() should call
following helper in their setup instead of the prior :

	dev->priv_flags &= ~IFF_XMIT_DST_RELEASE;
->
	netif_keep_dst(dev);

Instead of using a single bit, we use two bits, one being
eventually rebuilt in bonding/team drivers.

The other one, is permanent and blocks IFF_XMIT_DST_RELEASE being
rebuilt in bonding/team. Eventually, we could add something
smarter later.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Julian Anastasov <ja@ssi.bg>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-10-07 13:22:11 -04:00

608 lines
15 KiB
C
Raw Blame History

/*
* Equalizer Load-balancer for serial network interfaces.
*
* (c) Copyright 1995 Simon "Guru Aleph-Null" Janes
* NCM: Network and Communications Management, Inc.
*
* (c) Copyright 2002 David S. Miller (davem@redhat.com)
*
* This software may be used and distributed according to the terms
* of the GNU General Public License, incorporated herein by reference.
*
* The author may be reached as simon@ncm.com, or C/O
* NCM
* Attn: Simon Janes
* 6803 Whittier Ave
* McLean VA 22101
* Phone: 1-703-847-0040 ext 103
*/
/*
* Sources:
* skeleton.c by Donald Becker.
* Inspirations:
* The Harried and Overworked Alan Cox
* Conspiracies:
* The Alan Cox and Mike McLagan plot to get someone else to do the code,
* which turned out to be me.
*/
/*
* $Log: eql.c,v $
* Revision 1.2 1996/04/11 17:51:52 guru
* Added one-line eql_remove_slave patch.
*
* Revision 1.1 1996/04/11 17:44:17 guru
* Initial revision
*
* Revision 3.13 1996/01/21 15:17:18 alan
* tx_queue_len changes.
* reformatted.
*
* Revision 3.12 1995/03/22 21:07:51 anarchy
* Added capable() checks on configuration.
* Moved header file.
*
* Revision 3.11 1995/01/19 23:14:31 guru
* slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
* (priority_Bps) + bytes_queued * 8;
*
* Revision 3.10 1995/01/19 23:07:53 guru
* back to
* slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
* (priority_Bps) + bytes_queued;
*
* Revision 3.9 1995/01/19 22:38:20 guru
* slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
* (priority_Bps) + bytes_queued * 4;
*
* Revision 3.8 1995/01/19 22:30:55 guru
* slave_load = (ULONG_MAX - (ULONG_MAX / 2)) -
* (priority_Bps) + bytes_queued * 2;
*
* Revision 3.7 1995/01/19 21:52:35 guru
* printk's trimmed out.
*
* Revision 3.6 1995/01/19 21:49:56 guru
* This is working pretty well. I gained 1 K/s in speed.. now it's just
* robustness and printk's to be diked out.
*
* Revision 3.5 1995/01/18 22:29:59 guru
* still crashes the kernel when the lock_wait thing is woken up.
*
* Revision 3.4 1995/01/18 21:59:47 guru
* Broken set-bit locking snapshot
*
* Revision 3.3 1995/01/17 22:09:18 guru
* infinite sleep in a lock somewhere..
*
* Revision 3.2 1995/01/15 16:46:06 guru
* Log trimmed of non-pertinent 1.x branch messages
*
* Revision 3.1 1995/01/15 14:41:45 guru
* New Scheduler and timer stuff...
*
* Revision 1.15 1995/01/15 14:29:02 guru
* Will make 1.14 (now 1.15) the 3.0 branch, and the 1.12 the 2.0 branch, the one
* with the dumber scheduler
*
* Revision 1.14 1995/01/15 02:37:08 guru
* shock.. the kept-new-versions could have zonked working
* stuff.. shudder
*
* Revision 1.13 1995/01/15 02:36:31 guru
* big changes
*
* scheduler was torn out and replaced with something smarter
*
* global names not prefixed with eql_ were renamed to protect
* against namespace collisions
*
* a few more abstract interfaces were added to facilitate any
* potential change of datastructure. the driver is still using
* a linked list of slaves. going to a heap would be a bit of
* an overkill.
*
* this compiles fine with no warnings.
*
* the locking mechanism and timer stuff must be written however,
* this version will not work otherwise
*
* Sorry, I had to rewrite most of this for 2.5.x -DaveM
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/capability.h>
#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/slab.h>
#include <linux/timer.h>
#include <linux/netdevice.h>
#include <net/net_namespace.h>
#include <linux/if.h>
#include <linux/if_arp.h>
#include <linux/if_eql.h>
#include <linux/pkt_sched.h>
#include <asm/uaccess.h>
static int eql_open(struct net_device *dev);
static int eql_close(struct net_device *dev);
static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd);
static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev);
#define eql_is_slave(dev) ((dev->flags & IFF_SLAVE) == IFF_SLAVE)
#define eql_is_master(dev) ((dev->flags & IFF_MASTER) == IFF_MASTER)
static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave);
static void eql_timer(unsigned long param)
{
equalizer_t *eql = (equalizer_t *) param;
struct list_head *this, *tmp, *head;
spin_lock(&eql->queue.lock);
head = &eql->queue.all_slaves;
list_for_each_safe(this, tmp, head) {
slave_t *slave = list_entry(this, slave_t, list);
if ((slave->dev->flags & IFF_UP) == IFF_UP) {
slave->bytes_queued -= slave->priority_Bps;
if (slave->bytes_queued < 0)
slave->bytes_queued = 0;
} else {
eql_kill_one_slave(&eql->queue, slave);
}
}
spin_unlock(&eql->queue.lock);
eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL;
add_timer(&eql->timer);
}
static const char version[] __initconst =
"Equalizer2002: Simon Janes (simon@ncm.com) and David S. Miller (davem@redhat.com)";
static const struct net_device_ops eql_netdev_ops = {
.ndo_open = eql_open,
.ndo_stop = eql_close,
.ndo_do_ioctl = eql_ioctl,
.ndo_start_xmit = eql_slave_xmit,
};
static void __init eql_setup(struct net_device *dev)
{
equalizer_t *eql = netdev_priv(dev);
init_timer(&eql->timer);
eql->timer.data = (unsigned long) eql;
eql->timer.expires = jiffies + EQL_DEFAULT_RESCHED_IVAL;
eql->timer.function = eql_timer;
spin_lock_init(&eql->queue.lock);
INIT_LIST_HEAD(&eql->queue.all_slaves);
eql->queue.master_dev = dev;
dev->netdev_ops = &eql_netdev_ops;
/*
* Now we undo some of the things that eth_setup does
* that we don't like
*/
dev->mtu = EQL_DEFAULT_MTU; /* set to 576 in if_eql.h */
dev->flags = IFF_MASTER;
dev->type = ARPHRD_SLIP;
dev->tx_queue_len = 5; /* Hands them off fast */
netif_keep_dst(dev);
}
static int eql_open(struct net_device *dev)
{
equalizer_t *eql = netdev_priv(dev);
/* XXX We should force this off automatically for the user. */
netdev_info(dev,
"remember to turn off Van-Jacobson compression on your slave devices\n");
BUG_ON(!list_empty(&eql->queue.all_slaves));
eql->min_slaves = 1;
eql->max_slaves = EQL_DEFAULT_MAX_SLAVES; /* 4 usually... */
add_timer(&eql->timer);
return 0;
}
static void eql_kill_one_slave(slave_queue_t *queue, slave_t *slave)
{
list_del(&slave->list);
queue->num_slaves--;
slave->dev->flags &= ~IFF_SLAVE;
dev_put(slave->dev);
kfree(slave);
}
static void eql_kill_slave_queue(slave_queue_t *queue)
{
struct list_head *head, *tmp, *this;
spin_lock_bh(&queue->lock);
head = &queue->all_slaves;
list_for_each_safe(this, tmp, head) {
slave_t *s = list_entry(this, slave_t, list);
eql_kill_one_slave(queue, s);
}
spin_unlock_bh(&queue->lock);
}
static int eql_close(struct net_device *dev)
{
equalizer_t *eql = netdev_priv(dev);
/*
* The timer has to be stopped first before we start hacking away
* at the data structure it scans every so often...
*/
del_timer_sync(&eql->timer);
eql_kill_slave_queue(&eql->queue);
return 0;
}
static int eql_enslave(struct net_device *dev, slaving_request_t __user *srq);
static int eql_emancipate(struct net_device *dev, slaving_request_t __user *srq);
static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *sc);
static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mc);
static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mc);
static int eql_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
{
if (cmd != EQL_GETMASTRCFG && cmd != EQL_GETSLAVECFG &&
!capable(CAP_NET_ADMIN))
return -EPERM;
switch (cmd) {
case EQL_ENSLAVE:
return eql_enslave(dev, ifr->ifr_data);
case EQL_EMANCIPATE:
return eql_emancipate(dev, ifr->ifr_data);
case EQL_GETSLAVECFG:
return eql_g_slave_cfg(dev, ifr->ifr_data);
case EQL_SETSLAVECFG:
return eql_s_slave_cfg(dev, ifr->ifr_data);
case EQL_GETMASTRCFG:
return eql_g_master_cfg(dev, ifr->ifr_data);
case EQL_SETMASTRCFG:
return eql_s_master_cfg(dev, ifr->ifr_data);
default:
return -EOPNOTSUPP;
}
}
/* queue->lock must be held */
static slave_t *__eql_schedule_slaves(slave_queue_t *queue)
{
unsigned long best_load = ~0UL;
struct list_head *this, *tmp, *head;
slave_t *best_slave;
best_slave = NULL;
/* Make a pass to set the best slave. */
head = &queue->all_slaves;
list_for_each_safe(this, tmp, head) {
slave_t *slave = list_entry(this, slave_t, list);
unsigned long slave_load, bytes_queued, priority_Bps;
/* Go through the slave list once, updating best_slave
* whenever a new best_load is found.
*/
bytes_queued = slave->bytes_queued;
priority_Bps = slave->priority_Bps;
if ((slave->dev->flags & IFF_UP) == IFF_UP) {
slave_load = (~0UL - (~0UL / 2)) -
(priority_Bps) + bytes_queued * 8;
if (slave_load < best_load) {
best_load = slave_load;
best_slave = slave;
}
} else {
/* We found a dead slave, kill it. */
eql_kill_one_slave(queue, slave);
}
}
return best_slave;
}
static netdev_tx_t eql_slave_xmit(struct sk_buff *skb, struct net_device *dev)
{
equalizer_t *eql = netdev_priv(dev);
slave_t *slave;
spin_lock(&eql->queue.lock);
slave = __eql_schedule_slaves(&eql->queue);
if (slave) {
struct net_device *slave_dev = slave->dev;
skb->dev = slave_dev;
skb->priority = TC_PRIO_FILLER;
slave->bytes_queued += skb->len;
dev_queue_xmit(skb);
dev->stats.tx_packets++;
} else {
dev->stats.tx_dropped++;
dev_kfree_skb(skb);
}
spin_unlock(&eql->queue.lock);
return NETDEV_TX_OK;
}
/*
* Private ioctl functions
*/
/* queue->lock must be held */
static slave_t *__eql_find_slave_dev(slave_queue_t *queue, struct net_device *dev)
{
struct list_head *this, *head;
head = &queue->all_slaves;
list_for_each(this, head) {
slave_t *slave = list_entry(this, slave_t, list);
if (slave->dev == dev)
return slave;
}
return NULL;
}
static inline int eql_is_full(slave_queue_t *queue)
{
equalizer_t *eql = netdev_priv(queue->master_dev);
if (queue->num_slaves >= eql->max_slaves)
return 1;
return 0;
}
/* queue->lock must be held */
static int __eql_insert_slave(slave_queue_t *queue, slave_t *slave)
{
if (!eql_is_full(queue)) {
slave_t *duplicate_slave = NULL;
duplicate_slave = __eql_find_slave_dev(queue, slave->dev);
if (duplicate_slave)
eql_kill_one_slave(queue, duplicate_slave);
dev_hold(slave->dev);
list_add(&slave->list, &queue->all_slaves);
queue->num_slaves++;
slave->dev->flags |= IFF_SLAVE;
return 0;
}
return -ENOSPC;
}
static int eql_enslave(struct net_device *master_dev, slaving_request_t __user *srqp)
{
struct net_device *slave_dev;
slaving_request_t srq;
if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
return -EFAULT;
slave_dev = __dev_get_by_name(&init_net, srq.slave_name);
if (!slave_dev)
return -ENODEV;
if ((master_dev->flags & IFF_UP) == IFF_UP) {
/* slave is not a master & not already a slave: */
if (!eql_is_master(slave_dev) && !eql_is_slave(slave_dev)) {
slave_t *s = kmalloc(sizeof(*s), GFP_KERNEL);
equalizer_t *eql = netdev_priv(master_dev);
int ret;
if (!s)
return -ENOMEM;
memset(s, 0, sizeof(*s));
s->dev = slave_dev;
s->priority = srq.priority;
s->priority_bps = srq.priority;
s->priority_Bps = srq.priority / 8;
spin_lock_bh(&eql->queue.lock);
ret = __eql_insert_slave(&eql->queue, s);
if (ret)
kfree(s);
spin_unlock_bh(&eql->queue.lock);
return ret;
}
}
return -EINVAL;
}
static int eql_emancipate(struct net_device *master_dev, slaving_request_t __user *srqp)
{
equalizer_t *eql = netdev_priv(master_dev);
struct net_device *slave_dev;
slaving_request_t srq;
int ret;
if (copy_from_user(&srq, srqp, sizeof (slaving_request_t)))
return -EFAULT;
slave_dev = __dev_get_by_name(&init_net, srq.slave_name);
if (!slave_dev)
return -ENODEV;
ret = -EINVAL;
spin_lock_bh(&eql->queue.lock);
if (eql_is_slave(slave_dev)) {
slave_t *slave = __eql_find_slave_dev(&eql->queue, slave_dev);
if (slave) {
eql_kill_one_slave(&eql->queue, slave);
ret = 0;
}
}
spin_unlock_bh(&eql->queue.lock);
return ret;
}
static int eql_g_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
{
equalizer_t *eql = netdev_priv(dev);
slave_t *slave;
struct net_device *slave_dev;
slave_config_t sc;
int ret;
if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
return -EFAULT;
slave_dev = __dev_get_by_name(&init_net, sc.slave_name);
if (!slave_dev)
return -ENODEV;
ret = -EINVAL;
spin_lock_bh(&eql->queue.lock);
if (eql_is_slave(slave_dev)) {
slave = __eql_find_slave_dev(&eql->queue, slave_dev);
if (slave) {
sc.priority = slave->priority;
ret = 0;
}
}
spin_unlock_bh(&eql->queue.lock);
if (!ret && copy_to_user(scp, &sc, sizeof (slave_config_t)))
ret = -EFAULT;
return ret;
}
static int eql_s_slave_cfg(struct net_device *dev, slave_config_t __user *scp)
{
slave_t *slave;
equalizer_t *eql;
struct net_device *slave_dev;
slave_config_t sc;
int ret;
if (copy_from_user(&sc, scp, sizeof (slave_config_t)))
return -EFAULT;
slave_dev = __dev_get_by_name(&init_net, sc.slave_name);
if (!slave_dev)
return -ENODEV;
ret = -EINVAL;
eql = netdev_priv(dev);
spin_lock_bh(&eql->queue.lock);
if (eql_is_slave(slave_dev)) {
slave = __eql_find_slave_dev(&eql->queue, slave_dev);
if (slave) {
slave->priority = sc.priority;
slave->priority_bps = sc.priority;
slave->priority_Bps = sc.priority / 8;
ret = 0;
}
}
spin_unlock_bh(&eql->queue.lock);
return ret;
}
static int eql_g_master_cfg(struct net_device *dev, master_config_t __user *mcp)
{
equalizer_t *eql;
master_config_t mc;
memset(&mc, 0, sizeof(master_config_t));
if (eql_is_master(dev)) {
eql = netdev_priv(dev);
mc.max_slaves = eql->max_slaves;
mc.min_slaves = eql->min_slaves;
if (copy_to_user(mcp, &mc, sizeof (master_config_t)))
return -EFAULT;
return 0;
}
return -EINVAL;
}
static int eql_s_master_cfg(struct net_device *dev, master_config_t __user *mcp)
{
equalizer_t *eql;
master_config_t mc;
if (copy_from_user(&mc, mcp, sizeof (master_config_t)))
return -EFAULT;
if (eql_is_master(dev)) {
eql = netdev_priv(dev);
eql->max_slaves = mc.max_slaves;
eql->min_slaves = mc.min_slaves;
return 0;
}
return -EINVAL;
}
static struct net_device *dev_eql;
static int __init eql_init_module(void)
{
int err;
pr_info("%s\n", version);
dev_eql = alloc_netdev(sizeof(equalizer_t), "eql", NET_NAME_UNKNOWN,
eql_setup);
if (!dev_eql)
return -ENOMEM;
err = register_netdev(dev_eql);
if (err)
free_netdev(dev_eql);
return err;
}
static void __exit eql_cleanup_module(void)
{
unregister_netdev(dev_eql);
free_netdev(dev_eql);
}
module_init(eql_init_module);
module_exit(eql_cleanup_module);
MODULE_LICENSE("GPL");
Reference in New Issue View Git Blame Copy Permalink