[net-next-2.6.git] / drivers / net / cxgb3 / l2t.c

/*
 * Copyright (c) 2003-2008 Chelsio, Inc. All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#include <linux/skbuff.h>
#include <linux/netdevice.h>
#include <linux/if.h>
#include <linux/if_vlan.h>
#include <linux/jhash.h>
#include <linux/slab.h>
#include <net/neighbour.h>
#include "common.h"
#include "t3cdev.h"
#include "cxgb3_defs.h"
#include "l2t.h"
#include "t3_cpl.h"
#include "firmware_exports.h"

#define VLAN_NONE 0xfff

/*
 * Module locking notes:  There is a RW lock protecting the L2 table as a
 * whole plus a spinlock per L2T entry.  Entry lookups and allocations happen
 * under the protection of the table lock, individual entry changes happen
 * while holding that entry's spinlock.  The table lock nests outside the
 * entry locks.  Allocations of new entries take the table lock as writers so
 * no other lookups can happen while allocating new entries.  Entry updates
 * take the table lock as readers so multiple entries can be updated in
 * parallel.  An L2T entry can be dropped by decrementing its reference count
 * and therefore can happen in parallel with entry allocation but no entry
 * can change state or increment its ref count during allocation as both of
 * these perform lookups.
 */

static inline unsigned int vlan_prio(const struct l2t_entry *e)
{
	return e->vlan >> 13;
}

static inline unsigned int arp_hash(u32 key, int ifindex,
				    const struct l2t_data *d)
{
	return jhash_2words(key, ifindex, 0) & (d->nentries - 1);
}

static inline void neigh_replace(struct l2t_entry *e, struct neighbour *n)
{
	neigh_hold(n);
	if (e->neigh)
		neigh_release(e->neigh);
	e->neigh = n;
}

/*
 * Set up an L2T entry and send any packets waiting in the arp queue.  The
 * supplied skb is used for the CPL_L2T_WRITE_REQ.  Must be called with the
 * entry locked.
 */
static int setup_l2e_send_pending(struct t3cdev *dev, struct sk_buff *skb,
				  struct l2t_entry *e)
{
	struct cpl_l2t_write_req *req;
	struct sk_buff *tmp;

	if (!skb) {
		skb = alloc_skb(sizeof(*req), GFP_ATOMIC);
		if (!skb)
			return -ENOMEM;
	}

	req = (struct cpl_l2t_write_req *)__skb_put(skb, sizeof(*req));
	req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_FORWARD));
	OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_L2T_WRITE_REQ, e->idx));
	req->params = htonl(V_L2T_W_IDX(e->idx) | V_L2T_W_IFF(e->smt_idx) |
			    V_L2T_W_VLAN(e->vlan & VLAN_VID_MASK) |
			    V_L2T_W_PRIO(vlan_prio(e)));
	memcpy(e->dmac, e->neigh->ha, sizeof(e->dmac));
	memcpy(req->dst_mac, e->dmac, sizeof(req->dst_mac));
	skb->priority = CPL_PRIORITY_CONTROL;
	cxgb3_ofld_send(dev, skb);

	skb_queue_walk_safe(&e->arpq, skb, tmp) {
		__skb_unlink(skb, &e->arpq);
		cxgb3_ofld_send(dev, skb);
	}
	e->state = L2T_STATE_VALID;

	return 0;
}

/*
 * Add a packet to the an L2T entry's queue of packets awaiting resolution.
 * Must be called with the entry's lock held.
 */
static inline void arpq_enqueue(struct l2t_entry *e, struct sk_buff *skb)
{
	__skb_queue_tail(&e->arpq, skb);
}

int t3_l2t_send_slow(struct t3cdev *dev, struct sk_buff *skb,
		     struct l2t_entry *e)
{
again:
	switch (e->state) {
	case L2T_STATE_STALE:	/* entry is stale, kick off revalidation */
		neigh_event_send(e->neigh, NULL);
		spin_lock_bh(&e->lock);
		if (e->state == L2T_STATE_STALE)
			e->state = L2T_STATE_VALID;
		spin_unlock_bh(&e->lock);
	case L2T_STATE_VALID:	/* fast-path, send the packet on */
		return cxgb3_ofld_send(dev, skb);
	case L2T_STATE_RESOLVING:
		spin_lock_bh(&e->lock);
		if (e->state != L2T_STATE_RESOLVING) {
			/* ARP already completed */
			spin_unlock_bh(&e->lock);
			goto again;
		}
		arpq_enqueue(e, skb);
		spin_unlock_bh(&e->lock);

		/*
		 * Only the first packet added to the arpq should kick off
		 * resolution.  However, because the alloc_skb below can fail,
		 * we allow each packet added to the arpq to retry resolution
		 * as a way of recovering from transient memory exhaustion.
		 * A better way would be to use a work request to retry L2T
		 * entries when there's no memory.
		 */
		if (!neigh_event_send(e->neigh, NULL)) {
			skb = alloc_skb(sizeof(struct cpl_l2t_write_req),
					GFP_ATOMIC);
			if (!skb)
				break;

			spin_lock_bh(&e->lock);
			if (!skb_queue_empty(&e->arpq))
				setup_l2e_send_pending(dev, skb, e);
			else	/* we lost the race */
				__kfree_skb(skb);
			spin_unlock_bh(&e->lock);
		}
	}
	return 0;
}

EXPORT_SYMBOL(t3_l2t_send_slow);

void t3_l2t_send_event(struct t3cdev *dev, struct l2t_entry *e)
{
again:
	switch (e->state) {
	case L2T_STATE_STALE:	/* entry is stale, kick off revalidation */
		neigh_event_send(e->neigh, NULL);
		spin_lock_bh(&e->lock);
		if (e->state == L2T_STATE_STALE) {
			e->state = L2T_STATE_VALID;
		}
		spin_unlock_bh(&e->lock);
		return;
	case L2T_STATE_VALID:	/* fast-path, send the packet on */
		return;
	case L2T_STATE_RESOLVING:
		spin_lock_bh(&e->lock);
		if (e->state != L2T_STATE_RESOLVING) {
			/* ARP already completed */
			spin_unlock_bh(&e->lock);
			goto again;
		}
		spin_unlock_bh(&e->lock);

		/*
		 * Only the first packet added to the arpq should kick off
		 * resolution.  However, because the alloc_skb below can fail,
		 * we allow each packet added to the arpq to retry resolution
		 * as a way of recovering from transient memory exhaustion.
		 * A better way would be to use a work request to retry L2T
		 * entries when there's no memory.
		 */
		neigh_event_send(e->neigh, NULL);
	}
	return;
}

EXPORT_SYMBOL(t3_l2t_send_event);

/*
 * Allocate a free L2T entry.  Must be called with l2t_data.lock held.
 */
static struct l2t_entry *alloc_l2e(struct l2t_data *d)
{
	struct l2t_entry *end, *e, **p;

	if (!atomic_read(&d->nfree))
		return NULL;

	/* there's definitely a free entry */
	for (e = d->rover, end = &d->l2tab[d->nentries]; e != end; ++e)
		if (atomic_read(&e->refcnt) == 0)
			goto found;

	for (e = &d->l2tab[1]; atomic_read(&e->refcnt); ++e) ;
found:
	d->rover = e + 1;
	atomic_dec(&d->nfree);

	/*
	 * The entry we found may be an inactive entry that is
	 * presently in the hash table.  We need to remove it.
	 */
	if (e->state != L2T_STATE_UNUSED) {
		int hash = arp_hash(e->addr, e->ifindex, d);

		for (p = &d->l2tab[hash].first; *p; p = &(*p)->next)
			if (*p == e) {
				*p = e->next;
				break;
			}
		e->state = L2T_STATE_UNUSED;
	}
	return e;
}

/*
 * Called when an L2T entry has no more users.  The entry is left in the hash
 * table since it is likely to be reused but we also bump nfree to indicate
 * that the entry can be reallocated for a different neighbor.  We also drop
 * the existing neighbor reference in case the neighbor is going away and is
 * waiting on our reference.
 *
 * Because entries can be reallocated to other neighbors once their ref count
 * drops to 0 we need to take the entry's lock to avoid races with a new
 * incarnation.
 */
void t3_l2e_free(struct l2t_data *d, struct l2t_entry *e)
{
	spin_lock_bh(&e->lock);
	if (atomic_read(&e->refcnt) == 0) {	/* hasn't been recycled */
		if (e->neigh) {
			neigh_release(e->neigh);
			e->neigh = NULL;
		}
	}
	spin_unlock_bh(&e->lock);
	atomic_inc(&d->nfree);
}

EXPORT_SYMBOL(t3_l2e_free);

/*
 * Update an L2T entry that was previously used for the same next hop as neigh.
 * Must be called with softirqs disabled.
 */
static inline void reuse_entry(struct l2t_entry *e, struct neighbour *neigh)
{
	unsigned int nud_state;

	spin_lock(&e->lock);	/* avoid race with t3_l2t_free */

	if (neigh != e->neigh)
		neigh_replace(e, neigh);
	nud_state = neigh->nud_state;
	if (memcmp(e->dmac, neigh->ha, sizeof(e->dmac)) ||
	    !(nud_state & NUD_VALID))
		e->state = L2T_STATE_RESOLVING;
	else if (nud_state & NUD_CONNECTED)
		e->state = L2T_STATE_VALID;
	else
		e->state = L2T_STATE_STALE;
	spin_unlock(&e->lock);
}

struct l2t_entry *t3_l2t_get(struct t3cdev *cdev, struct neighbour *neigh,
			     struct net_device *dev)
{
	struct l2t_entry *e;
	struct l2t_data *d = L2DATA(cdev);
	u32 addr = *(u32 *) neigh->primary_key;
	int ifidx = neigh->dev->ifindex;
	int hash = arp_hash(addr, ifidx, d);
	struct port_info *p = netdev_priv(dev);
	int smt_idx = p->port_id;

	write_lock_bh(&d->lock);
	for (e = d->l2tab[hash].first; e; e = e->next)
		if (e->addr == addr && e->ifindex == ifidx &&
		    e->smt_idx == smt_idx) {
			l2t_hold(d, e);
			if (atomic_read(&e->refcnt) == 1)
				reuse_entry(e, neigh);
			goto done;
		}

	/* Need to allocate a new entry */
	e = alloc_l2e(d);
	if (e) {
		spin_lock(&e->lock);	/* avoid race with t3_l2t_free */
		e->next = d->l2tab[hash].first;
		d->l2tab[hash].first = e;
		e->state = L2T_STATE_RESOLVING;
		e->addr = addr;
		e->ifindex = ifidx;
		e->smt_idx = smt_idx;
		atomic_set(&e->refcnt, 1);
		neigh_replace(e, neigh);
		if (neigh->dev->priv_flags & IFF_802_1Q_VLAN)
			e->vlan = vlan_dev_vlan_id(neigh->dev);
		else
			e->vlan = VLAN_NONE;
		spin_unlock(&e->lock);
	}
done:
	write_unlock_bh(&d->lock);
	return e;
}

EXPORT_SYMBOL(t3_l2t_get);

/*
 * Called when address resolution fails for an L2T entry to handle packets
 * on the arpq head.  If a packet specifies a failure handler it is invoked,
 * otherwise the packets is sent to the offload device.
 *
 * XXX: maybe we should abandon the latter behavior and just require a failure
 * handler.
 */
static void handle_failed_resolution(struct t3cdev *dev, struct sk_buff_head *arpq)
{
	struct sk_buff *skb, *tmp;

	skb_queue_walk_safe(arpq, skb, tmp) {
		struct l2t_skb_cb *cb = L2T_SKB_CB(skb);

		__skb_unlink(skb, arpq);
		if (cb->arp_failure_handler)
			cb->arp_failure_handler(dev, skb);
		else
			cxgb3_ofld_send(dev, skb);
	}
}

/*
 * Called when the host's ARP layer makes a change to some entry that is
 * loaded into the HW L2 table.
 */
void t3_l2t_update(struct t3cdev *dev, struct neighbour *neigh)
{
	struct sk_buff_head arpq;
	struct l2t_entry *e;
	struct l2t_data *d = L2DATA(dev);
	u32 addr = *(u32 *) neigh->primary_key;
	int ifidx = neigh->dev->ifindex;
	int hash = arp_hash(addr, ifidx, d);

	read_lock_bh(&d->lock);
	for (e = d->l2tab[hash].first; e; e = e->next)
		if (e->addr == addr && e->ifindex == ifidx) {
			spin_lock(&e->lock);
			goto found;
		}
	read_unlock_bh(&d->lock);
	return;

found:
	__skb_queue_head_init(&arpq);

	read_unlock(&d->lock);
	if (atomic_read(&e->refcnt)) {
		if (neigh != e->neigh)
			neigh_replace(e, neigh);

		if (e->state == L2T_STATE_RESOLVING) {
			if (neigh->nud_state & NUD_FAILED) {
				skb_queue_splice_init(&e->arpq, &arpq);
			} else if (neigh->nud_state & (NUD_CONNECTED|NUD_STALE))
				setup_l2e_send_pending(dev, NULL, e);
		} else {
			e->state = neigh->nud_state & NUD_CONNECTED ?
			    L2T_STATE_VALID : L2T_STATE_STALE;
			if (memcmp(e->dmac, neigh->ha, 6))
				setup_l2e_send_pending(dev, NULL, e);
		}
	}
	spin_unlock_bh(&e->lock);

	if (!skb_queue_empty(&arpq))
		handle_failed_resolution(dev, &arpq);
}

struct l2t_data *t3_init_l2t(unsigned int l2t_capacity)
{
	struct l2t_data *d;
	int i, size = sizeof(*d) + l2t_capacity * sizeof(struct l2t_entry);

	d = cxgb_alloc_mem(size);
	if (!d)
		return NULL;

	d->nentries = l2t_capacity;
	d->rover = &d->l2tab[1];	/* entry 0 is not used */
	atomic_set(&d->nfree, l2t_capacity - 1);
	rwlock_init(&d->lock);

	for (i = 0; i < l2t_capacity; ++i) {
		d->l2tab[i].idx = i;
		d->l2tab[i].state = L2T_STATE_UNUSED;
		__skb_queue_head_init(&d->l2tab[i].arpq);
		spin_lock_init(&d->l2tab[i].lock);
		atomic_set(&d->l2tab[i].refcnt, 0);
	}
	return d;
}

void t3_free_l2t(struct l2t_data *d)
{
	cxgb_free_mem(d);
}
Commit	Line	Data
4d22de3e	1	/*
a02d44a0	2	* Copyright (c) 2003-2008 Chelsio, Inc. All rights reserved.
4d22de3e DLR	3	*
	4	* This software is available to you under a choice of one of two
	5	* licenses. You may choose to be licensed under the terms of the GNU
	6	* General Public License (GPL) Version 2, available from the file
	7	* COPYING in the main directory of this source tree, or the
	8	* OpenIB.org BSD license below:
	9	*
	10	* Redistribution and use in source and binary forms, with or
	11	* without modification, are permitted provided that the following
	12	* conditions are met:
	13	*
	14	* - Redistributions of source code must retain the above
	15	* copyright notice, this list of conditions and the following
	16	* disclaimer.
	17	*
	18	* - Redistributions in binary form must reproduce the above
	19	* copyright notice, this list of conditions and the following
	20	* disclaimer in the documentation and/or other materials
	21	* provided with the distribution.
	22	*
	23	* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
	24	* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
	25	* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
	26	* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
	27	* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
	28	* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
	29	* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
	30	* SOFTWARE.
	31	*/
	32	#include <linux/skbuff.h>
	33	#include <linux/netdevice.h>
	34	#include <linux/if.h>
	35	#include <linux/if_vlan.h>
	36	#include <linux/jhash.h>
5a0e3ad6	37	#include <linux/slab.h>
4d22de3e DLR	38	#include <net/neighbour.h>
	39	#include "common.h"
	40	#include "t3cdev.h"
	41	#include "cxgb3_defs.h"
	42	#include "l2t.h"
	43	#include "t3_cpl.h"
	44	#include "firmware_exports.h"
	45
	46	#define VLAN_NONE 0xfff
	47
	48	/*
	49	* Module locking notes: There is a RW lock protecting the L2 table as a
	50	* whole plus a spinlock per L2T entry. Entry lookups and allocations happen
	51	* under the protection of the table lock, individual entry changes happen
	52	* while holding that entry's spinlock. The table lock nests outside the
	53	* entry locks. Allocations of new entries take the table lock as writers so
	54	* no other lookups can happen while allocating new entries. Entry updates
	55	* take the table lock as readers so multiple entries can be updated in
	56	* parallel. An L2T entry can be dropped by decrementing its reference count
	57	* and therefore can happen in parallel with entry allocation but no entry
	58	* can change state or increment its ref count during allocation as both of
	59	* these perform lookups.
	60	*/
	61
	62	static inline unsigned int vlan_prio(const struct l2t_entry *e)
	63	{
	64	return e->vlan >> 13;
	65	}
	66
	67	static inline unsigned int arp_hash(u32 key, int ifindex,
	68	const struct l2t_data *d)
	69	{
	70	return jhash_2words(key, ifindex, 0) & (d->nentries - 1);
	71	}
	72
	73	static inline void neigh_replace(struct l2t_entry e, struct neighbour n)
	74	{
	75	neigh_hold(n);
	76	if (e->neigh)
	77	neigh_release(e->neigh);
	78	e->neigh = n;
	79	}
	80
	81	/*
	82	* Set up an L2T entry and send any packets waiting in the arp queue. The
	83	* supplied skb is used for the CPL_L2T_WRITE_REQ. Must be called with the
	84	* entry locked.
	85	*/
	86	static int setup_l2e_send_pending(struct t3cdev dev, struct sk_buff skb,
	87	struct l2t_entry *e)
	88	{
	89	struct cpl_l2t_write_req *req;
147e70e6	90	struct sk_buff *tmp;
4d22de3e DLR	91
	92	if (!skb) {
	93	skb = alloc_skb(sizeof(*req), GFP_ATOMIC);
	94	if (!skb)
	95	return -ENOMEM;
	96	}
	97
	98	req = (struct cpl_l2t_write_req )__skb_put(skb, sizeof(req));
	99	req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_FORWARD));
	100	OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_L2T_WRITE_REQ, e->idx));
	101	req->params = htonl(V_L2T_W_IDX(e->idx) \| V_L2T_W_IFF(e->smt_idx) \|
	102	V_L2T_W_VLAN(e->vlan & VLAN_VID_MASK) \|
	103	V_L2T_W_PRIO(vlan_prio(e)));
	104	memcpy(e->dmac, e->neigh->ha, sizeof(e->dmac));
	105	memcpy(req->dst_mac, e->dmac, sizeof(req->dst_mac));
	106	skb->priority = CPL_PRIORITY_CONTROL;
	107	cxgb3_ofld_send(dev, skb);
147e70e6 DM	108
	109	skb_queue_walk_safe(&e->arpq, skb, tmp) {
	110	__skb_unlink(skb, &e->arpq);
4d22de3e DLR	111	cxgb3_ofld_send(dev, skb);
4d22de3e DLR	112	}
4d22de3e DLR	113	e->state = L2T_STATE_VALID;
	114
	115	return 0;
	116	}
	117
	118	/*
	119	* Add a packet to the an L2T entry's queue of packets awaiting resolution.
	120	* Must be called with the entry's lock held.
	121	*/
	122	static inline void arpq_enqueue(struct l2t_entry e, struct sk_buff skb)
	123	{
147e70e6	124	__skb_queue_tail(&e->arpq, skb);
4d22de3e DLR	125	}
	126
	127	int t3_l2t_send_slow(struct t3cdev dev, struct sk_buff skb,
	128	struct l2t_entry *e)
	129	{
	130	again:
	131	switch (e->state) {
	132	case L2T_STATE_STALE: /* entry is stale, kick off revalidation */
	133	neigh_event_send(e->neigh, NULL);
	134	spin_lock_bh(&e->lock);
	135	if (e->state == L2T_STATE_STALE)
	136	e->state = L2T_STATE_VALID;
	137	spin_unlock_bh(&e->lock);
	138	case L2T_STATE_VALID: /* fast-path, send the packet on */
	139	return cxgb3_ofld_send(dev, skb);
	140	case L2T_STATE_RESOLVING:
	141	spin_lock_bh(&e->lock);
	142	if (e->state != L2T_STATE_RESOLVING) {
	143	/* ARP already completed */
	144	spin_unlock_bh(&e->lock);
	145	goto again;
	146	}
	147	arpq_enqueue(e, skb);
	148	spin_unlock_bh(&e->lock);
	149
	150	/*
	151	* Only the first packet added to the arpq should kick off
	152	* resolution. However, because the alloc_skb below can fail,
	153	* we allow each packet added to the arpq to retry resolution
	154	* as a way of recovering from transient memory exhaustion.
	155	* A better way would be to use a work request to retry L2T
	156	* entries when there's no memory.
	157	*/
	158	if (!neigh_event_send(e->neigh, NULL)) {
	159	skb = alloc_skb(sizeof(struct cpl_l2t_write_req),
	160	GFP_ATOMIC);
	161	if (!skb)
	162	break;
	163
	164	spin_lock_bh(&e->lock);
147e70e6	165	if (!skb_queue_empty(&e->arpq))
4d22de3e DLR	166	setup_l2e_send_pending(dev, skb, e);
	167	else /* we lost the race */
	168	__kfree_skb(skb);
	169	spin_unlock_bh(&e->lock);
	170	}
	171	}
	172	return 0;
	173	}
	174
	175	EXPORT_SYMBOL(t3_l2t_send_slow);
	176
	177	void t3_l2t_send_event(struct t3cdev dev, struct l2t_entry e)
	178	{
	179	again:
	180	switch (e->state) {
	181	case L2T_STATE_STALE: /* entry is stale, kick off revalidation */
	182	neigh_event_send(e->neigh, NULL);
	183	spin_lock_bh(&e->lock);
	184	if (e->state == L2T_STATE_STALE) {
	185	e->state = L2T_STATE_VALID;
	186	}
	187	spin_unlock_bh(&e->lock);
	188	return;
	189	case L2T_STATE_VALID: /* fast-path, send the packet on */
	190	return;
	191	case L2T_STATE_RESOLVING:
	192	spin_lock_bh(&e->lock);
	193	if (e->state != L2T_STATE_RESOLVING) {
	194	/* ARP already completed */
	195	spin_unlock_bh(&e->lock);
	196	goto again;
	197	}
	198	spin_unlock_bh(&e->lock);
	199
	200	/*
	201	* Only the first packet added to the arpq should kick off
	202	* resolution. However, because the alloc_skb below can fail,
	203	* we allow each packet added to the arpq to retry resolution
	204	* as a way of recovering from transient memory exhaustion.
	205	* A better way would be to use a work request to retry L2T
	206	* entries when there's no memory.
	207	*/
	208	neigh_event_send(e->neigh, NULL);
	209	}
	210	return;
	211	}
	212
	213	EXPORT_SYMBOL(t3_l2t_send_event);
	214
	215	/*
	216	* Allocate a free L2T entry. Must be called with l2t_data.lock held.
	217	*/
	218	static struct l2t_entry alloc_l2e(struct l2t_data d)
	219	{
	220	struct l2t_entry end, e, **p;
	221
	222	if (!atomic_read(&d->nfree))
	223	return NULL;
	224
	225	/* there's definitely a free entry */
	226	for (e = d->rover, end = &d->l2tab[d->nentries]; e != end; ++e)
	227	if (atomic_read(&e->refcnt) == 0)
	228	goto found;
	229
230	for (e = &d->l2tab[1]; atomic_read(&e->refcnt); ++e) ;
231	found:
232	d->rover = e + 1;
233	atomic_dec(&d->nfree);
234
235	/*
236	* The entry we found may be an inactive entry that is
237	* presently in the hash table. We need to remove it.
238	*/
239	if (e->state != L2T_STATE_UNUSED) {
240	int hash = arp_hash(e->addr, e->ifindex, d);
241
242	for (p = &d->l2tab[hash].first; p; p = &(p)->next)
243	if (*p == e) {
244	*p = e->next;
245	break;
246	}
247	e->state = L2T_STATE_UNUSED;
248	}
249	return e;
250	}
251
252	/*
253	* Called when an L2T entry has no more users. The entry is left in the hash
254	* table since it is likely to be reused but we also bump nfree to indicate
255	* that the entry can be reallocated for a different neighbor. We also drop
256	* the existing neighbor reference in case the neighbor is going away and is
257	* waiting on our reference.
258	*
259	* Because entries can be reallocated to other neighbors once their ref count
260	* drops to 0 we need to take the entry's lock to avoid races with a new
261	* incarnation.
262	*/
263	void t3_l2e_free(struct l2t_data d, struct l2t_entry e)
264	{
265	spin_lock_bh(&e->lock);
266	if (atomic_read(&e->refcnt) == 0) { /* hasn't been recycled */
267	if (e->neigh) {
268	neigh_release(e->neigh);
269	e->neigh = NULL;
270	}
271	}
272	spin_unlock_bh(&e->lock);
273	atomic_inc(&d->nfree);
274	}
275
276	EXPORT_SYMBOL(t3_l2e_free);
277
278	/*
279	* Update an L2T entry that was previously used for the same next hop as neigh.
280	* Must be called with softirqs disabled.
281	*/
282	static inline void reuse_entry(struct l2t_entry e, struct neighbour neigh)
283	{
284	unsigned int nud_state;
285
286	spin_lock(&e->lock); /* avoid race with t3_l2t_free */
287
288	if (neigh != e->neigh)
289	neigh_replace(e, neigh);
290	nud_state = neigh->nud_state;
291	if (memcmp(e->dmac, neigh->ha, sizeof(e->dmac)) \|\|
292	!(nud_state & NUD_VALID))
293	e->state = L2T_STATE_RESOLVING;
294	else if (nud_state & NUD_CONNECTED)
295	e->state = L2T_STATE_VALID;
296	else
297	e->state = L2T_STATE_STALE;
298	spin_unlock(&e->lock);
299	}
300
301	struct l2t_entry t3_l2t_get(struct t3cdev cdev, struct neighbour *neigh,
302	struct net_device *dev)
303	{
304	struct l2t_entry *e;
305	struct l2t_data *d = L2DATA(cdev);
306	u32 addr = (u32 ) neigh->primary_key;
307	int ifidx = neigh->dev->ifindex;
308	int hash = arp_hash(addr, ifidx, d);
309	struct port_info *p = netdev_priv(dev);
310	int smt_idx = p->port_id;
311
312	write_lock_bh(&d->lock);
313	for (e = d->l2tab[hash].first; e; e = e->next)
314	if (e->addr == addr && e->ifindex == ifidx &&
315	e->smt_idx == smt_idx) {
316	l2t_hold(d, e);
317	if (atomic_read(&e->refcnt) == 1)
318	reuse_entry(e, neigh);
319	goto done;
320	}
321
322	/* Need to allocate a new entry */
323	e = alloc_l2e(d);
324	if (e) {
325	spin_lock(&e->lock); /* avoid race with t3_l2t_free */
326	e->next = d->l2tab[hash].first;
327	d->l2tab[hash].first = e;
328	e->state = L2T_STATE_RESOLVING;
329	e->addr = addr;
330	e->ifindex = ifidx;
331	e->smt_idx = smt_idx;
332	atomic_set(&e->refcnt, 1);
333	neigh_replace(e, neigh);
334	if (neigh->dev->priv_flags & IFF_802_1Q_VLAN)
22d1ba74	335	e->vlan = vlan_dev_vlan_id(neigh->dev);
4d22de3e DLR	336	else
	337	e->vlan = VLAN_NONE;
	338	spin_unlock(&e->lock);
	339	}
	340	done:
	341	write_unlock_bh(&d->lock);
	342	return e;
	343	}
	344
	345	EXPORT_SYMBOL(t3_l2t_get);
	346
	347	/*
	348	* Called when address resolution fails for an L2T entry to handle packets
	349	* on the arpq head. If a packet specifies a failure handler it is invoked,
	350	* otherwise the packets is sent to the offload device.
	351	*
	352	* XXX: maybe we should abandon the latter behavior and just require a failure
	353	* handler.
	354	*/
147e70e6	355	static void handle_failed_resolution(struct t3cdev dev, struct sk_buff_head arpq)
4d22de3e	356	{
147e70e6 DM	357	struct sk_buff skb, tmp;
	358
	359	skb_queue_walk_safe(arpq, skb, tmp) {
4d22de3e DLR	360	struct l2t_skb_cb *cb = L2T_SKB_CB(skb);
4d22de3e DLR	361
147e70e6	362	__skb_unlink(skb, arpq);
4d22de3e DLR	363	if (cb->arp_failure_handler)
	364	cb->arp_failure_handler(dev, skb);
	365	else
	366	cxgb3_ofld_send(dev, skb);
	367	}
	368	}
	369
	370	/*
	371	* Called when the host's ARP layer makes a change to some entry that is
	372	* loaded into the HW L2 table.
	373	*/
	374	void t3_l2t_update(struct t3cdev dev, struct neighbour neigh)
	375	{
147e70e6	376	struct sk_buff_head arpq;
4d22de3e	377	struct l2t_entry *e;
4d22de3e DLR	378	struct l2t_data *d = L2DATA(dev);
	379	u32 addr = (u32 ) neigh->primary_key;
	380	int ifidx = neigh->dev->ifindex;
	381	int hash = arp_hash(addr, ifidx, d);
	382
	383	read_lock_bh(&d->lock);
	384	for (e = d->l2tab[hash].first; e; e = e->next)
	385	if (e->addr == addr && e->ifindex == ifidx) {
	386	spin_lock(&e->lock);
	387	goto found;
	388	}
	389	read_unlock_bh(&d->lock);
	390	return;
	391
	392	found:
147e70e6 DM	393	__skb_queue_head_init(&arpq);
147e70e6 DM	394
4d22de3e DLR	395	read_unlock(&d->lock);
	396	if (atomic_read(&e->refcnt)) {
	397	if (neigh != e->neigh)
	398	neigh_replace(e, neigh);
	399
	400	if (e->state == L2T_STATE_RESOLVING) {
	401	if (neigh->nud_state & NUD_FAILED) {
147e70e6	402	skb_queue_splice_init(&e->arpq, &arpq);
4eb61e02	403	} else if (neigh->nud_state & (NUD_CONNECTED\|NUD_STALE))
4d22de3e DLR	404	setup_l2e_send_pending(dev, NULL, e);
4d22de3e DLR	405	} else {
8082c37c	406	e->state = neigh->nud_state & NUD_CONNECTED ?
4d22de3e DLR	407	L2T_STATE_VALID : L2T_STATE_STALE;
	408	if (memcmp(e->dmac, neigh->ha, 6))
	409	setup_l2e_send_pending(dev, NULL, e);
	410	}
	411	}
	412	spin_unlock_bh(&e->lock);
	413
147e70e6 DM	414	if (!skb_queue_empty(&arpq))
147e70e6 DM	415	handle_failed_resolution(dev, &arpq);
4d22de3e DLR	416	}
	417
	418	struct l2t_data *t3_init_l2t(unsigned int l2t_capacity)
	419	{
	420	struct l2t_data *d;
	421	int i, size = sizeof(d) + l2t_capacity sizeof(struct l2t_entry);
	422
	423	d = cxgb_alloc_mem(size);
	424	if (!d)
	425	return NULL;
	426
	427	d->nentries = l2t_capacity;
	428	d->rover = &d->l2tab[1]; /* entry 0 is not used */
	429	atomic_set(&d->nfree, l2t_capacity - 1);
	430	rwlock_init(&d->lock);
	431
	432	for (i = 0; i < l2t_capacity; ++i) {
	433	d->l2tab[i].idx = i;
	434	d->l2tab[i].state = L2T_STATE_UNUSED;
6d329af9	435	__skb_queue_head_init(&d->l2tab[i].arpq);
4d22de3e DLR	436	spin_lock_init(&d->l2tab[i].lock);
	437	atomic_set(&d->l2tab[i].refcnt, 0);
	438	}
	439	return d;
	440	}
	441
	442	void t3_free_l2t(struct l2t_data *d)
	443	{
	444	cxgb_free_mem(d);
	445	}
	446