X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=kernel%2Fvserver%2Fnetwork.c;h=4c3eb435a52135f4b88b62be9a63f9dfe9fc980a;hb=6a77f38946aaee1cd85eeec6cf4229b204c15071;hp=479a19b47af0b0595e5e41775138ee8da38f1442;hpb=b76fcd5f0c655b6e3e9bf534594357025421c66a;p=linux-2.6.git

diff --git a/kernel/vserver/network.c b/kernel/vserver/network.c
index 479a19b47..4c3eb435a 100644
--- a/kernel/vserver/network.c
+++ b/kernel/vserver/network.c
@@ -3,160 +3,216 @@
  *
  *  Virtual Server: Network Support
  *
- *  Copyright (C) 2003-2004  Herbert Pötzl
+ *  Copyright (C) 2003-2005  Herbert Pötzl
  *
  *  V0.01  broken out from vcontext V0.05
  *  V0.02  cleaned up implementation
  *  V0.03  added equiv nx commands
+ *  V0.04  switch to RCU based hash
  *
  */
 
 #include <linux/config.h>
 #include <linux/slab.h>
-#include <linux/vserver/network.h>
-#include <linux/ninline.h>
+#include <linux/vserver/network_cmd.h>
+#include <linux/rcupdate.h>
+#include <net/tcp.h>
 
 #include <asm/errno.h>
 
 
-LIST_HEAD(nx_infos);
+/*	__alloc_nx_info()
 
-spinlock_t nxlist_lock
-	__cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
+	* allocate an initialized nx_info struct
+	* doesn't make it visible (hash)			*/
 
-
-/*
- *	struct nx_info allocation and deallocation
- */
-
-static struct nx_info *alloc_nx_info(void)
+static struct nx_info *__alloc_nx_info(nid_t nid)
 {
 	struct nx_info *new = NULL;
-	
-	nxdprintk("alloc_nx_info()\n");
+
+	vxdprintk(VXD_CBIT(nid, 1), "alloc_nx_info(%d)*", nid);
+
 	/* would this benefit from a slab cache? */
 	new = kmalloc(sizeof(struct nx_info), GFP_KERNEL);
 	if (!new)
 		return 0;
-	
+
 	memset (new, 0, sizeof(struct nx_info));
+	new->nx_id = nid;
+	INIT_RCU_HEAD(&new->nx_rcu);
+	INIT_HLIST_NODE(&new->nx_hlist);
+	atomic_set(&new->nx_refcnt, 0);
+	atomic_set(&new->nx_usecnt, 0);
+
 	/* rest of init goes here */
-	
-	nxdprintk("alloc_nx_info() = %p\n", new);
+
+	vxdprintk(VXD_CBIT(nid, 0),
+		"alloc_nx_info() = %p", new);
 	return new;
 }
 
-void free_nx_info(struct nx_info *nxi)
+/*	__dealloc_nx_info()
+
+	* final disposal of nx_info				*/
+
+static void __dealloc_nx_info(struct nx_info *nxi)
 {
-	nxdprintk("free_nx_info(%p)\n", nxi);
+	vxdprintk(VXD_CBIT(nid, 0),
+		"dealloc_nx_info(%p)", nxi);
+
+	nxi->nx_hlist.next = LIST_POISON1;
+	nxi->nx_id = -1;
+
+	BUG_ON(atomic_read(&nxi->nx_usecnt));
+	BUG_ON(atomic_read(&nxi->nx_refcnt));
+
 	kfree(nxi);
 }
 
-struct nx_info *create_nx_info(void)
+static inline int __free_nx_info(struct nx_info *nxi)
 {
-	struct nx_info *new;
-	static int gnid = 1;
-	
-	nxdprintk("create_nx_info()\n");
-	if (!(new = alloc_nx_info()))
-		return 0;
+	int usecnt, refcnt;
 
-	spin_lock(&nxlist_lock);
+	BUG_ON(!nxi);
 
-	/* new ip info */
-	atomic_set(&new->nx_refcount, 1);
-	new->nx_id = gnid++;
-	list_add(&new->nx_list, &nx_infos);
+	usecnt = atomic_read(&nxi->nx_usecnt);
+	BUG_ON(usecnt < 0);
 
-	spin_unlock(&nxlist_lock);
-	return new;
+	refcnt = atomic_read(&nxi->nx_refcnt);
+	BUG_ON(refcnt < 0);
+
+	if (!usecnt)
+		__dealloc_nx_info(nxi);
+	return usecnt;
 }
 
+/*	exported stuff						*/
+
+void free_nx_info(struct nx_info *nxi)
+{
+	/* context shutdown is mandatory */
+	// BUG_ON(nxi->nx_state != NXS_SHUTDOWN);
+
+	// BUG_ON(nxi->nx_state & NXS_HASHED);
+
+	BUG_ON(__free_nx_info(nxi));
+}
 
-/*
- *	struct nx_info search by id
- *	assumes nxlist_lock is held
- */
 
-static __inline__ struct nx_info *__find_nx_info(int id)
+/*	hash table for nx_info hash */
+
+#define NX_HASH_SIZE	13
+
+struct hlist_head nx_info_hash[NX_HASH_SIZE];
+
+static spinlock_t nx_info_hash_lock = SPIN_LOCK_UNLOCKED;
+
+
+static inline unsigned int __hashval(nid_t nid)
 {
-	struct nx_info *nxi;
+	return (nid % NX_HASH_SIZE);
+}
 
-	list_for_each_entry(nxi, &nx_infos, nx_list)
-		if (nxi->nx_id == id)
-			return nxi;
-	return 0;
+
+
+/*	__hash_nx_info()
+
+	* add the nxi to the global hash table
+	* requires the hash_lock to be held			*/
+
+static inline void __hash_nx_info(struct nx_info *nxi)
+{
+	struct hlist_head *head;
+
+	vxdprintk(VXD_CBIT(nid, 4),
+		"__hash_nx_info: %p[#%d]", nxi, nxi->nx_id);
+	get_nx_info(nxi);
+	head = &nx_info_hash[__hashval(nxi->nx_id)];
+	hlist_add_head(&nxi->nx_hlist, head);
 }
 
+/*	__unhash_nx_info()
 
-/*
- *	struct nx_info ref stuff
- */
+	* remove the nxi from the global hash table
+	* requires the hash_lock to be held			*/
 
-struct nx_info *find_nx_info(int id)
+static inline void __unhash_nx_info(struct nx_info *nxi)
 {
-	struct nx_info *nxi;
-	
-	if (id < 0) {
-		nxi = current->nx_info;
-		get_nx_info(nxi);
-	} else {
-		spin_lock(&nxlist_lock);
-		if ((nxi = __find_nx_info(id)))
-			get_nx_info(nxi);
-		spin_unlock(&nxlist_lock);
-	}
-	return nxi;
+	vxd_assert_lock(&nx_info_hash_lock);
+	vxdprintk(VXD_CBIT(nid, 4),
+		"__unhash_nx_info: %p[#%d]", nxi, nxi->nx_id);
+	hlist_del(&nxi->nx_hlist);
+	put_nx_info(nxi);
 }
 
-/*
- *      verify that id is a valid nid
- */
 
-int nx_info_id_valid(int id)
+/*	__lookup_nx_info()
+
+	* requires the hash_lock to be held
+	* doesn't increment the nx_refcnt			*/
+
+static inline struct nx_info *__lookup_nx_info(nid_t nid)
 {
-	int valid;
-	
-	spin_lock(&nxlist_lock);
-	valid = (__find_nx_info(id) != NULL);
-	spin_unlock(&nxlist_lock);
-	return valid;
+	struct hlist_head *head = &nx_info_hash[__hashval(nid)];
+	struct hlist_node *pos;
+
+	vxd_assert_lock(&nx_info_hash_lock);
+	hlist_for_each(pos, head) {
+		struct nx_info *nxi =
+			hlist_entry(pos, struct nx_info, nx_hlist);
+
+		if (nxi->nx_id == nid) {
+			return nxi;
+		}
+	}
+	return NULL;
 }
 
 
-/*
- *	dynamic context id ...
- */
+/*	__nx_dynamic_id()
+
+	* find unused dynamic nid
+	* requires the hash_lock to be held			*/
 
-static __inline__ nid_t __nx_dynamic_id(void)
+static inline nid_t __nx_dynamic_id(void)
 {
 	static nid_t seq = MAX_N_CONTEXT;
 	nid_t barrier = seq;
-	
+
+	vxd_assert_lock(&nx_info_hash_lock);
 	do {
 		if (++seq > MAX_N_CONTEXT)
 			seq = MIN_D_CONTEXT;
-		if (!__find_nx_info(seq))
+		if (!__lookup_nx_info(seq)) {
+			vxdprintk(VXD_CBIT(nid, 4),
+				"__nx_dynamic_id: [#%d]", seq);
 			return seq;
+		}
 	} while (barrier != seq);
 	return 0;
 }
 
-static struct nx_info * __foc_nx_info(int id, int *err)
+/*	__loc_nx_info()
+
+	* locate or create the requested context
+	* get() it and if new hash it				*/
+
+static struct nx_info * __loc_nx_info(int id, int *err)
 {
 	struct nx_info *new, *nxi = NULL;
-	
-	nxdprintk("foc_nx_info(%d)\n", id);
-	// if (!(new = alloc_nx_info(id))) {
-	if (!(new = alloc_nx_info())) {
+
+	vxdprintk(VXD_CBIT(nid, 1), "loc_nx_info(%d)*", id);
+
+	if (!(new = __alloc_nx_info(id))) {
 		*err = -ENOMEM;
 		return NULL;
 	}
 
-	spin_lock(&nxlist_lock);
+	/* required to make dynamic xids unique */
+	spin_lock(&nx_info_hash_lock);
 
 	/* dynamic context requested */
-	if (id == IP_DYNAMIC_ID) {
+	if (id == NX_DYNAMIC_ID) {
 		id = __nx_dynamic_id();
 		if (!id) {
 			printk(KERN_ERR "no dynamic context available.\n");
@@ -165,14 +221,16 @@ static struct nx_info * __foc_nx_info(int id, int *err)
 		new->nx_id = id;
 	}
 	/* existing context requested */
-	else if ((nxi = __find_nx_info(id))) {
+	else if ((nxi = __lookup_nx_info(id))) {
 		/* context in setup is not available */
 		if (nxi->nx_flags & VXF_STATE_SETUP) {
-			nxdprintk("foc_nx_info(%d) = %p (not available)\n", id, nxi);
+			vxdprintk(VXD_CBIT(nid, 0),
+				"loc_nx_info(%d) = %p (not available)", id, nxi);
 			nxi = NULL;
 			*err = -EBUSY;
 		} else {
-			nxdprintk("foc_nx_info(%d) = %p (found)\n", id, nxi);
+			vxdprintk(VXD_CBIT(nid, 0),
+				"loc_nx_info(%d) = %p (found)", id, nxi);
 			get_nx_info(nxi);
 			*err = 0;
 		}
@@ -180,50 +238,152 @@ static struct nx_info * __foc_nx_info(int id, int *err)
 	}
 
 	/* new context requested */
-	nxdprintk("foc_nx_info(%d) = %p (new)\n", id, new);
-	atomic_set(&new->nx_refcount, 1);
-	list_add(&new->nx_list, &nx_infos);
+	vxdprintk(VXD_CBIT(nid, 0),
+		"loc_nx_info(%d) = %p (new)", id, new);
+	__hash_nx_info(get_nx_info(new));
 	nxi = new, new = NULL;
 	*err = 1;
 
 out_unlock:
-	spin_unlock(&nxlist_lock);
+	spin_unlock(&nx_info_hash_lock);
 	if (new)
-		free_nx_info(new);
+		__dealloc_nx_info(new);
+	return nxi;
+}
+
+
+
+/*	exported stuff						*/
+
+
+void unhash_nx_info(struct nx_info *nxi)
+{
+	spin_lock(&nx_info_hash_lock);
+	__unhash_nx_info(nxi);
+	spin_unlock(&nx_info_hash_lock);
+}
+
+/*	locate_nx_info()
+
+	* search for a nx_info and get() it
+	* negative id means current				*/
+
+struct nx_info *locate_nx_info(int id)
+{
+	struct nx_info *nxi;
+
+	if (id < 0) {
+		nxi = get_nx_info(current->nx_info);
+	} else {
+		spin_lock(&nx_info_hash_lock);
+		nxi = get_nx_info(__lookup_nx_info(id));
+		spin_unlock(&nx_info_hash_lock);
+	}
 	return nxi;
 }
 
+/*	nid_is_hashed()
 
-struct nx_info *find_or_create_nx_info(int id)
+	* verify that nid is still hashed			*/
+
+int nid_is_hashed(nid_t nid)
 {
+	int hashed;
+
+	spin_lock(&nx_info_hash_lock);
+	hashed = (__lookup_nx_info(nid) != NULL);
+	spin_unlock(&nx_info_hash_lock);
+	return hashed;
+}
+
+#ifdef	CONFIG_VSERVER_LEGACYNET
+
+struct nx_info *locate_or_create_nx_info(int id)
+{
+	int err;
+
+	return __loc_nx_info(id, &err);
+}
+
+struct nx_info *create_nx_info(void)
+{
+	struct nx_info *new;
 	int err;
 
-	return __foc_nx_info(id, &err);
+	vxdprintk(VXD_CBIT(nid, 5), "create_nx_info(%s)", "void");
+	if (!(new = __loc_nx_info(NX_DYNAMIC_ID, &err)))
+		return NULL;
+	return new;
 }
 
+
+#endif
+
+#ifdef	CONFIG_PROC_FS
+
+int get_nid_list(int index, unsigned int *nids, int size)
+{
+	int hindex, nr_nids = 0;
+
+	for (hindex = 0; hindex < NX_HASH_SIZE; hindex++) {
+		struct hlist_head *head = &nx_info_hash[hindex];
+		struct hlist_node *pos;
+
+		spin_lock(&nx_info_hash_lock);
+		hlist_for_each(pos, head) {
+			struct nx_info *nxi;
+
+			if (--index > 0)
+				continue;
+
+			nxi = hlist_entry(pos, struct nx_info, nx_hlist);
+			nids[nr_nids] = nxi->nx_id;
+			if (++nr_nids >= size) {
+				spin_unlock(&nx_info_hash_lock);
+				goto out;
+			}
+		}
+		/* keep the lock time short */
+		spin_unlock(&nx_info_hash_lock);
+	}
+out:
+	return nr_nids;
+}
+#endif
+
+
 /*
  *	migrate task to new network
  */
 
 int nx_migrate_task(struct task_struct *p, struct nx_info *nxi)
 {
-	struct nx_info *old_nxi = task_get_nx_info(p);
+	struct nx_info *old_nxi;
 	int ret = 0;
-	
+
 	if (!p || !nxi)
 		BUG();
 
-	nxdprintk("nx_migrate_task(%p,%p[#%d.%d)\n", p, nxi,
-		nxi->nx_id, atomic_read(&nxi->nx_refcount));
+	vxdprintk(VXD_CBIT(nid, 5),
+		"nx_migrate_task(%p,%p[#%d.%d.%d])",
+		p, nxi, nxi->nx_id,
+		atomic_read(&nxi->nx_usecnt),
+		atomic_read(&nxi->nx_refcnt));
+
+	old_nxi = task_get_nx_info(p);
 	if (old_nxi == nxi)
 		goto out;
 
 	task_lock(p);
+	/* should be handled in set_nx_info !! */
+	if (old_nxi)
+		clr_nx_info(&p->nx_info);
 	set_nx_info(&p->nx_info, nxi);
 	p->nid = nxi->nx_id;
 	task_unlock(p);
 
-	put_nx_info(old_nxi);
+	/* obsoleted by clr/set */
+	// put_nx_info(old_nxi);
 out:
 	put_nx_info(old_nxi);
 	return ret;
@@ -233,23 +393,14 @@ out:
 #include <linux/netdevice.h>
 #include <linux/inetdevice.h>
 
-static inline int __addr_in_nx_info(u32 addr, struct nx_info *nxi)
-{
-	int i, nbip;
-
-	nbip = nxi->nbipv4;
-	for (i=0; i<nbip; i++)
-		if (nxi->ipv4[i] == addr)
-			return 1;
-	return 0;
-}
 
 int ifa_in_nx_info(struct in_ifaddr *ifa, struct nx_info *nxi)
 {
 	if (!nxi)
 		return 1;
-	
-	return __addr_in_nx_info(ifa->ifa_address, nxi);
+	if (!ifa)
+		return 0;
+	return addr_in_nx_info(nxi, ifa->ifa_address);
 }
 
 int dev_in_nx_info(struct net_device *dev, struct nx_info *nxi)
@@ -265,13 +416,63 @@ int dev_in_nx_info(struct net_device *dev, struct nx_info *nxi)
 
 	for (ifap = &in_dev->ifa_list; (ifa = *ifap) != NULL;
 		ifap = &ifa->ifa_next) {
-		if (__addr_in_nx_info(ifa->ifa_address, nxi))
+		if (addr_in_nx_info(nxi, ifa->ifa_address))
 			return 1;
 	}
 	return 0;
 }
 
+/*
+ *	check if address is covered by socket
+ *
+ *	sk:	the socket to check against
+ *	addr:	the address in question (must be != 0)
+ */
+static inline int __addr_in_socket(struct sock *sk, uint32_t addr)
+{
+	struct nx_info *nxi = sk->sk_nx_info;
+	uint32_t saddr = tcp_v4_rcv_saddr(sk);
+
+	vxdprintk(VXD_CBIT(net, 5),
+		"__addr_in_socket(%p,%d.%d.%d.%d) %p:%d.%d.%d.%d %p;%lx",
+		sk, VXD_QUAD(addr), nxi, VXD_QUAD(saddr), sk->sk_socket,
+		(sk->sk_socket?sk->sk_socket->flags:0));
+
+	if (saddr) {
+		/* direct address match */
+		return (saddr == addr);
+	} else if (nxi) {
+		/* match against nx_info */
+		return addr_in_nx_info(nxi, addr);
+	} else {
+		/* unrestricted any socket */
+		return 1;
+	}
+}
+
 
+int nx_addr_conflict(struct nx_info *nxi, uint32_t addr, struct sock *sk)
+{
+	vxdprintk(VXD_CBIT(net, 2),
+		"nx_addr_conflict(%p,%p) %d.%d,%d.%d",
+		nxi, sk, VXD_QUAD(addr));
+
+	if (addr) {
+		/* check real address */
+		return __addr_in_socket(sk, addr);
+	} else if (nxi) {
+		/* check against nx_info */
+		int i, n = nxi->nbipv4;
+
+		for (i=0; i<n; i++)
+			if (__addr_in_socket(sk, nxi->ipv4[i]))
+				return 1;
+		return 0;
+	} else {
+		/* check against any */
+		return 1;
+	}
+}
 
 
 /* vserver syscall commands below here */
@@ -283,22 +484,22 @@ int dev_in_nx_info(struct net_device *dev, struct nx_info *nxi)
 
 int vc_task_nid(uint32_t id, void __user *data)
 {
-        nid_t nid;
-
-        if (id) {
-                struct task_struct *tsk;
-
-                if (!vx_check(0, VX_ADMIN|VX_WATCH))
-                        return -EPERM;
-
-                read_lock(&tasklist_lock);
-                tsk = find_task_by_pid(id);
-                nid = (tsk) ? tsk->nid : -ESRCH;
-                read_unlock(&tasklist_lock);
-        }
-        else
-                nid = current->nid;
-        return nid;
+	nid_t nid;
+
+	if (id) {
+		struct task_struct *tsk;
+
+		if (!vx_check(0, VX_ADMIN|VX_WATCH))
+			return -EPERM;
+
+		read_lock(&tasklist_lock);
+		tsk = find_task_by_real_pid(id);
+		nid = (tsk) ? tsk->nid : -ESRCH;
+		read_unlock(&tasklist_lock);
+	}
+	else
+		nid = current->nid;
+	return nid;
 }
 
 
@@ -312,7 +513,7 @@ int vc_nx_info(uint32_t id, void __user *data)
 	if (!capable(CAP_SYS_ADMIN) || !capable(CAP_SYS_RESOURCE))
 		return -EPERM;
 
-	nxi = find_nx_info(id);
+	nxi = locate_nx_info(id);
 	if (!nxi)
 		return -ESRCH;
 
@@ -329,7 +530,7 @@ int vc_nx_info(uint32_t id, void __user *data)
 
 int vc_net_create(uint32_t nid, void __user *data)
 {
-        // int ret = -ENOMEM;
+	// int ret = -ENOMEM;
 	struct nx_info *new_nxi;
 	int ret;
 
@@ -342,7 +543,7 @@ int vc_net_create(uint32_t nid, void __user *data)
 	if (nid < 1)
 		return -EINVAL;
 
-	new_nxi = __foc_nx_info(nid, &ret);
+	new_nxi = __loc_nx_info(nid, &ret);
 	if (!new_nxi)
 		return ret;
 	if (!(new_nxi->nx_flags & VXF_STATE_SETUP)) {
@@ -361,11 +562,11 @@ out_put:
 int vc_net_migrate(uint32_t id, void __user *data)
 {
 	struct nx_info *nxi;
-	
+
 	if (!capable(CAP_SYS_ADMIN))
 		return -EPERM;
 
-	nxi = find_nx_info(id);
+	nxi = locate_nx_info(id);
 	if (!nxi)
 		return -ESRCH;
 	nx_migrate_task(current, nxi);
@@ -383,7 +584,7 @@ int vc_net_add(uint32_t id, void __user *data)
 	if (copy_from_user (&vc_data, data, sizeof(vc_data)))
 		return -EFAULT;
 
-	nxi = find_nx_info(id);
+	nxi = locate_nx_info(id);
 	if (!nxi)
 		return -ESRCH;
 
@@ -402,7 +603,7 @@ int vc_net_remove(uint32_t id, void __user *data)
 	if (copy_from_user (&vc_data, data, sizeof(vc_data)))
 		return -EFAULT;
 
-	nxi = find_nx_info(id);
+	nxi = locate_nx_info(id);
 	if (!nxi)
 		return -ESRCH;
 
@@ -421,13 +622,12 @@ int vc_get_nflags(uint32_t id, void __user *data)
 	if (!capable(CAP_SYS_ADMIN))
 		return -EPERM;
 
-	nxi = find_nx_info(id);
+	nxi = locate_nx_info(id);
 	if (!nxi)
 		return -ESRCH;
 
 	vc_data.flagword = nxi->nx_flags;
 
-	// vc_data.mask = ~0UL;
 	/* special STATE flag handling */
 	vc_data.mask = vx_mask_flags(~0UL, nxi->nx_flags, IPF_ONE_TIME);
 
@@ -449,7 +649,7 @@ int vc_set_nflags(uint32_t id, void __user *data)
 	if (copy_from_user (&vc_data, data, sizeof(vc_data)))
 		return -EFAULT;
 
-	nxi = find_nx_info(id);
+	nxi = locate_nx_info(id);
 	if (!nxi)
 		return -ESRCH;
 
@@ -472,7 +672,7 @@ int vc_get_ncaps(uint32_t id, void __user *data)
 	if (!capable(CAP_SYS_ADMIN))
 		return -EPERM;
 
-	nxi = find_nx_info(id);
+	nxi = locate_nx_info(id);
 	if (!nxi)
 		return -ESRCH;
 
@@ -495,7 +695,7 @@ int vc_set_ncaps(uint32_t id, void __user *data)
 	if (copy_from_user (&vc_data, data, sizeof(vc_data)))
 		return -EFAULT;
 
-	nxi = find_nx_info(id);
+	nxi = locate_nx_info(id);
 	if (!nxi)
 		return -ESRCH;
 
@@ -509,5 +709,5 @@ int vc_set_ncaps(uint32_t id, void __user *data)
 #include <linux/module.h>
 
 EXPORT_SYMBOL_GPL(free_nx_info);
-EXPORT_SYMBOL_GPL(nxlist_lock);
+EXPORT_SYMBOL_GPL(unhash_nx_info);