freebsd-dev/sys/netgraph/ng_base.c

3627 lines
94 KiB
C
Raw Normal View History

/*
* ng_base.c
*/
/*-
* Copyright (c) 1996-1999 Whistle Communications, Inc.
* All rights reserved.
*
* Subject to the following obligations and disclaimer of warranty, use and
* redistribution of this software, in source or object code forms, with or
* without modifications are expressly permitted by Whistle Communications;
* provided, however, that:
* 1. Any and all reproductions of the source or object code must include the
* copyright notice above and the following disclaimer of warranties; and
* 2. No rights are granted, in any manner or form, to use Whistle
* Communications, Inc. trademarks, including the mark "WHISTLE
* COMMUNICATIONS" on advertising, endorsements, or otherwise except as
* such appears in the above copyright notice or in the software.
*
* THIS SOFTWARE IS BEING PROVIDED BY WHISTLE COMMUNICATIONS "AS IS", AND
* TO THE MAXIMUM EXTENT PERMITTED BY LAW, WHISTLE COMMUNICATIONS MAKES NO
* REPRESENTATIONS OR WARRANTIES, EXPRESS OR IMPLIED, REGARDING THIS SOFTWARE,
* INCLUDING WITHOUT LIMITATION, ANY AND ALL IMPLIED WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, OR NON-INFRINGEMENT.
* WHISTLE COMMUNICATIONS DOES NOT WARRANT, GUARANTEE, OR MAKE ANY
* REPRESENTATIONS REGARDING THE USE OF, OR THE RESULTS OF THE USE OF THIS
* SOFTWARE IN TERMS OF ITS CORRECTNESS, ACCURACY, RELIABILITY OR OTHERWISE.
* IN NO EVENT SHALL WHISTLE COMMUNICATIONS BE LIABLE FOR ANY DAMAGES
* RESULTING FROM OR ARISING OUT OF ANY USE OF THIS SOFTWARE, INCLUDING
* WITHOUT LIMITATION, ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY,
* PUNITIVE, OR CONSEQUENTIAL DAMAGES, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES, LOSS OF USE, DATA OR PROFITS, HOWEVER CAUSED AND UNDER ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
* THIS SOFTWARE, EVEN IF WHISTLE COMMUNICATIONS IS ADVISED OF THE POSSIBILITY
* OF SUCH DAMAGE.
*
* Authors: Julian Elischer <julian@freebsd.org>
* Archie Cobbs <archie@freebsd.org>
*
* $FreeBSD$
* $Whistle: ng_base.c,v 1.39 1999/01/28 23:54:53 julian Exp $
*/
/*
* This file implements the base netgraph code.
*/
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/ctype.h>
#include <sys/errno.h>
#include <sys/kdb.h>
#include <sys/kernel.h>
#include <sys/limits.h>
#include <sys/malloc.h>
#include <sys/mbuf.h>
#include <sys/queue.h>
#include <sys/sysctl.h>
#include <sys/syslog.h>
#include <net/netisr.h>
#include <netgraph/ng_message.h>
#include <netgraph/netgraph.h>
#include <netgraph/ng_parse.h>
MODULE_VERSION(netgraph, NG_ABI_VERSION);
/* List of all active nodes */
static LIST_HEAD(, ng_node) ng_nodelist;
static struct mtx ng_nodelist_mtx;
#ifdef NETGRAPH_DEBUG
static struct mtx ngq_mtx; /* protects the queue item list */
static SLIST_HEAD(, ng_node) ng_allnodes;
static LIST_HEAD(, ng_node) ng_freenodes; /* in debug, we never free() them */
static SLIST_HEAD(, ng_hook) ng_allhooks;
static LIST_HEAD(, ng_hook) ng_freehooks; /* in debug, we never free() them */
static void ng_dumpitems(void);
static void ng_dumpnodes(void);
static void ng_dumphooks(void);
#endif /* NETGRAPH_DEBUG */
/*
* DEAD versions of the structures.
* In order to avoid races, it is sometimes neccesary to point
* at SOMETHING even though theoretically, the current entity is
* INVALID. Use these to avoid these races.
*/
struct ng_type ng_deadtype = {
NG_ABI_VERSION,
"dead",
NULL, /* modevent */
NULL, /* constructor */
NULL, /* rcvmsg */
NULL, /* shutdown */
NULL, /* newhook */
NULL, /* findhook */
NULL, /* connect */
NULL, /* rcvdata */
NULL, /* disconnect */
NULL, /* cmdlist */
};
struct ng_node ng_deadnode = {
"dead",
&ng_deadtype,
NGF_INVALID,
1, /* refs */
0, /* numhooks */
NULL, /* private */
0, /* ID */
LIST_HEAD_INITIALIZER(ng_deadnode.hooks),
{}, /* all_nodes list entry */
{}, /* id hashtable list entry */
{}, /* workqueue entry */
{ 0,
{}, /* should never use! (should hang) */
NULL,
&ng_deadnode.nd_input_queue.queue,
&ng_deadnode
},
#ifdef NETGRAPH_DEBUG
ND_MAGIC,
__FILE__,
__LINE__,
{NULL}
#endif /* NETGRAPH_DEBUG */
};
struct ng_hook ng_deadhook = {
"dead",
NULL, /* private */
HK_INVALID | HK_DEAD,
1, /* refs always >= 1 */
0, /* undefined data link type */
&ng_deadhook, /* Peer is self */
&ng_deadnode, /* attached to deadnode */
{}, /* hooks list */
NULL, /* override rcvmsg() */
NULL, /* override rcvdata() */
#ifdef NETGRAPH_DEBUG
HK_MAGIC,
__FILE__,
__LINE__,
{NULL}
#endif /* NETGRAPH_DEBUG */
};
/*
* END DEAD STRUCTURES
*/
/* List nodes with unallocated work */
static TAILQ_HEAD(, ng_node) ng_worklist = TAILQ_HEAD_INITIALIZER(ng_worklist);
static struct mtx ng_worklist_mtx; /* MUST LOCK NODE FIRST */
/* List of installed types */
static LIST_HEAD(, ng_type) ng_typelist;
static struct mtx ng_typelist_mtx;
/* Hash related definitions */
/* XXX Don't need to initialise them because it's a LIST */
#define NG_ID_HASH_SIZE 32 /* most systems wont need even this many */
static LIST_HEAD(, ng_node) ng_ID_hash[NG_ID_HASH_SIZE];
static struct mtx ng_idhash_mtx;
/* Method to find a node.. used twice so do it here */
#define NG_IDHASH_FN(ID) ((ID) % (NG_ID_HASH_SIZE))
#define NG_IDHASH_FIND(ID, node) \
do { \
mtx_assert(&ng_idhash_mtx, MA_OWNED); \
LIST_FOREACH(node, &ng_ID_hash[NG_IDHASH_FN(ID)], \
nd_idnodes) { \
if (NG_NODE_IS_VALID(node) \
&& (NG_NODE_ID(node) == ID)) { \
break; \
} \
} \
} while (0)
/* Internal functions */
static int ng_add_hook(node_p node, const char *name, hook_p * hookp);
static int ng_generic_msg(node_p here, item_p item, hook_p lasthook);
static ng_ID_t ng_decodeidname(const char *name);
static int ngb_mod_event(module_t mod, int event, void *data);
static void ng_worklist_remove(node_p node);
static void ngintr(void);
static int ng_apply_item(node_p node, item_p item);
static void ng_flush_input_queue(struct ng_queue * ngq);
static void ng_setisr(node_p node);
static node_p ng_ID2noderef(ng_ID_t ID);
static int ng_con_nodes(node_p node, const char *name, node_p node2,
const char *name2);
static void ng_con_part2(node_p node, hook_p hook, void *arg1, int arg2);
static void ng_con_part3(node_p node, hook_p hook, void *arg1, int arg2);
static int ng_mkpeer(node_p node, const char *name,
const char *name2, char *type);
/* imported , these used to be externally visible, some may go back */
void ng_destroy_hook(hook_p hook);
node_p ng_name2noderef(node_p node, const char *name);
int ng_path2noderef(node_p here, const char *path,
node_p *dest, hook_p *lasthook);
int ng_make_node(const char *type, node_p *nodepp);
int ng_path_parse(char *addr, char **node, char **path, char **hook);
void ng_rmnode(node_p node, hook_p dummy1, void *dummy2, int dummy3);
void ng_unname(node_p node);
/* Our own netgraph malloc type */
MALLOC_DEFINE(M_NETGRAPH, "netgraph", "netgraph structures and ctrl messages");
MALLOC_DEFINE(M_NETGRAPH_HOOK, "netgraph_hook", "netgraph hook structures");
MALLOC_DEFINE(M_NETGRAPH_NODE, "netgraph_node", "netgraph node structures");
MALLOC_DEFINE(M_NETGRAPH_ITEM, "netgraph_item", "netgraph item structures");
MALLOC_DEFINE(M_NETGRAPH_MSG, "netgraph_msg", "netgraph name storage");
/* Should not be visible outside this file */
#define _NG_ALLOC_HOOK(hook) \
MALLOC(hook, hook_p, sizeof(*hook), M_NETGRAPH_HOOK, M_NOWAIT | M_ZERO)
#define _NG_ALLOC_NODE(node) \
MALLOC(node, node_p, sizeof(*node), M_NETGRAPH_NODE, M_NOWAIT | M_ZERO)
#ifdef NETGRAPH_DEBUG /*----------------------------------------------*/
/*
* In debug mode:
* In an attempt to help track reference count screwups
* we do not free objects back to the malloc system, but keep them
* in a local cache where we can examine them and keep information safely
* after they have been freed.
* We use this scheme for nodes and hooks, and to some extent for items.
*/
static __inline hook_p
ng_alloc_hook(void)
{
hook_p hook;
SLIST_ENTRY(ng_hook) temp;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx);
hook = LIST_FIRST(&ng_freehooks);
if (hook) {
LIST_REMOVE(hook, hk_hooks);
bcopy(&hook->hk_all, &temp, sizeof(temp));
bzero(hook, sizeof(struct ng_hook));
bcopy(&temp, &hook->hk_all, sizeof(temp));
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
hook->hk_magic = HK_MAGIC;
} else {
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
_NG_ALLOC_HOOK(hook);
if (hook) {
hook->hk_magic = HK_MAGIC;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx);
SLIST_INSERT_HEAD(&ng_allhooks, hook, hk_all);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
}
}
return (hook);
}
static __inline node_p
ng_alloc_node(void)
{
node_p node;
SLIST_ENTRY(ng_node) temp;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx);
node = LIST_FIRST(&ng_freenodes);
if (node) {
LIST_REMOVE(node, nd_nodes);
bcopy(&node->nd_all, &temp, sizeof(temp));
bzero(node, sizeof(struct ng_node));
bcopy(&temp, &node->nd_all, sizeof(temp));
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
node->nd_magic = ND_MAGIC;
} else {
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
_NG_ALLOC_NODE(node);
if (node) {
node->nd_magic = ND_MAGIC;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx);
SLIST_INSERT_HEAD(&ng_allnodes, node, nd_all);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
}
}
return (node);
}
#define NG_ALLOC_HOOK(hook) do { (hook) = ng_alloc_hook(); } while (0)
#define NG_ALLOC_NODE(node) do { (node) = ng_alloc_node(); } while (0)
#define NG_FREE_HOOK(hook) \
do { \
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx); \
LIST_INSERT_HEAD(&ng_freehooks, hook, hk_hooks); \
hook->hk_magic = 0; \
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx); \
} while (0)
#define NG_FREE_NODE(node) \
do { \
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx); \
LIST_INSERT_HEAD(&ng_freenodes, node, nd_nodes); \
node->nd_magic = 0; \
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx); \
} while (0)
#else /* NETGRAPH_DEBUG */ /*----------------------------------------------*/
#define NG_ALLOC_HOOK(hook) _NG_ALLOC_HOOK(hook)
#define NG_ALLOC_NODE(node) _NG_ALLOC_NODE(node)
#define NG_FREE_HOOK(hook) do { FREE((hook), M_NETGRAPH_HOOK); } while (0)
#define NG_FREE_NODE(node) do { FREE((node), M_NETGRAPH_NODE); } while (0)
#endif /* NETGRAPH_DEBUG */ /*----------------------------------------------*/
/* Set this to kdb_enter("X") to catch all errors as they occur */
#ifndef TRAP_ERROR
#define TRAP_ERROR()
#endif
static ng_ID_t nextID = 1;
#ifdef INVARIANTS
#define CHECK_DATA_MBUF(m) do { \
struct mbuf *n; \
int total; \
\
M_ASSERTPKTHDR(m); \
for (total = 0, n = (m); n != NULL; n = n->m_next) \
total += n->m_len; \
if ((m)->m_pkthdr.len != total) { \
panic("%s: %d != %d", \
__func__, (m)->m_pkthdr.len, total); \
} \
} while (0)
#else
#define CHECK_DATA_MBUF(m)
#endif
/************************************************************************
Parse type definitions for generic messages
************************************************************************/
/* Handy structure parse type defining macro */
#define DEFINE_PARSE_STRUCT_TYPE(lo, up, args) \
static const struct ng_parse_struct_field \
ng_ ## lo ## _type_fields[] = NG_GENERIC_ ## up ## _INFO args; \
static const struct ng_parse_type ng_generic_ ## lo ## _type = { \
&ng_parse_struct_type, \
&ng_ ## lo ## _type_fields \
}
DEFINE_PARSE_STRUCT_TYPE(mkpeer, MKPEER, ());
DEFINE_PARSE_STRUCT_TYPE(connect, CONNECT, ());
DEFINE_PARSE_STRUCT_TYPE(name, NAME, ());
DEFINE_PARSE_STRUCT_TYPE(rmhook, RMHOOK, ());
DEFINE_PARSE_STRUCT_TYPE(nodeinfo, NODEINFO, ());
DEFINE_PARSE_STRUCT_TYPE(typeinfo, TYPEINFO, ());
DEFINE_PARSE_STRUCT_TYPE(linkinfo, LINKINFO, (&ng_generic_nodeinfo_type));
/* Get length of an array when the length is stored as a 32 bit
value immediately preceding the array -- as with struct namelist
and struct typelist. */
static int
ng_generic_list_getLength(const struct ng_parse_type *type,
const u_char *start, const u_char *buf)
{
return *((const u_int32_t *)(buf - 4));
}
/* Get length of the array of struct linkinfo inside a struct hooklist */
static int
ng_generic_linkinfo_getLength(const struct ng_parse_type *type,
const u_char *start, const u_char *buf)
{
const struct hooklist *hl = (const struct hooklist *)start;
return hl->nodeinfo.hooks;
}
/* Array type for a variable length array of struct namelist */
static const struct ng_parse_array_info ng_nodeinfoarray_type_info = {
&ng_generic_nodeinfo_type,
&ng_generic_list_getLength
};
static const struct ng_parse_type ng_generic_nodeinfoarray_type = {
&ng_parse_array_type,
&ng_nodeinfoarray_type_info
};
/* Array type for a variable length array of struct typelist */
static const struct ng_parse_array_info ng_typeinfoarray_type_info = {
&ng_generic_typeinfo_type,
&ng_generic_list_getLength
};
static const struct ng_parse_type ng_generic_typeinfoarray_type = {
&ng_parse_array_type,
&ng_typeinfoarray_type_info
};
/* Array type for array of struct linkinfo in struct hooklist */
static const struct ng_parse_array_info ng_generic_linkinfo_array_type_info = {
&ng_generic_linkinfo_type,
&ng_generic_linkinfo_getLength
};
static const struct ng_parse_type ng_generic_linkinfo_array_type = {
&ng_parse_array_type,
&ng_generic_linkinfo_array_type_info
};
DEFINE_PARSE_STRUCT_TYPE(typelist, TYPELIST, (&ng_generic_nodeinfoarray_type));
DEFINE_PARSE_STRUCT_TYPE(hooklist, HOOKLIST,
(&ng_generic_nodeinfo_type, &ng_generic_linkinfo_array_type));
DEFINE_PARSE_STRUCT_TYPE(listnodes, LISTNODES,
(&ng_generic_nodeinfoarray_type));
/* List of commands and how to convert arguments to/from ASCII */
static const struct ng_cmdlist ng_generic_cmds[] = {
{
NGM_GENERIC_COOKIE,
NGM_SHUTDOWN,
"shutdown",
NULL,
NULL
},
{
NGM_GENERIC_COOKIE,
NGM_MKPEER,
"mkpeer",
&ng_generic_mkpeer_type,
NULL
},
{
NGM_GENERIC_COOKIE,
NGM_CONNECT,
"connect",
&ng_generic_connect_type,
NULL
},
{
NGM_GENERIC_COOKIE,
NGM_NAME,
"name",
&ng_generic_name_type,
NULL
},
{
NGM_GENERIC_COOKIE,
NGM_RMHOOK,
"rmhook",
&ng_generic_rmhook_type,
NULL
},
{
NGM_GENERIC_COOKIE,
NGM_NODEINFO,
"nodeinfo",
NULL,
&ng_generic_nodeinfo_type
},
{
NGM_GENERIC_COOKIE,
NGM_LISTHOOKS,
"listhooks",
NULL,
&ng_generic_hooklist_type
},
{
NGM_GENERIC_COOKIE,
NGM_LISTNAMES,
"listnames",
NULL,
&ng_generic_listnodes_type /* same as NGM_LISTNODES */
},
{
NGM_GENERIC_COOKIE,
NGM_LISTNODES,
"listnodes",
NULL,
&ng_generic_listnodes_type
},
{
NGM_GENERIC_COOKIE,
NGM_LISTTYPES,
"listtypes",
NULL,
&ng_generic_typeinfo_type
},
{
NGM_GENERIC_COOKIE,
NGM_TEXT_CONFIG,
"textconfig",
NULL,
&ng_parse_string_type
},
{
NGM_GENERIC_COOKIE,
NGM_TEXT_STATUS,
"textstatus",
NULL,
&ng_parse_string_type
},
{
NGM_GENERIC_COOKIE,
NGM_ASCII2BINARY,
"ascii2binary",
&ng_parse_ng_mesg_type,
&ng_parse_ng_mesg_type
},
{
NGM_GENERIC_COOKIE,
NGM_BINARY2ASCII,
"binary2ascii",
&ng_parse_ng_mesg_type,
&ng_parse_ng_mesg_type
},
{ 0 }
};
/************************************************************************
Node routines
************************************************************************/
/*
* Instantiate a node of the requested type
*/
int
ng_make_node(const char *typename, node_p *nodepp)
{
struct ng_type *type;
int error;
/* Check that the type makes sense */
if (typename == NULL) {
TRAP_ERROR();
return (EINVAL);
}
/* Locate the node type. If we fail we return. Do not try to load
* module.
*/
if ((type = ng_findtype(typename)) == NULL)
return (ENXIO);
/*
* If we have a constructor, then make the node and
* call the constructor to do type specific initialisation.
*/
if (type->constructor != NULL) {
if ((error = ng_make_node_common(type, nodepp)) == 0) {
if ((error = ((*type->constructor)(*nodepp)) != 0)) {
NG_NODE_UNREF(*nodepp);
}
}
} else {
/*
* Node has no constructor. We cannot ask for one
* to be made. It must be brought into existance by
* some external agency. The external agency should
* call ng_make_node_common() directly to get the
* netgraph part initialised.
*/
TRAP_ERROR();
error = EINVAL;
}
return (error);
}
/*
* Generic node creation. Called by node initialisation for externally
* instantiated nodes (e.g. hardware, sockets, etc ).
* The returned node has a reference count of 1.
*/
int
ng_make_node_common(struct ng_type *type, node_p *nodepp)
{
node_p node;
/* Require the node type to have been already installed */
if (ng_findtype(type->name) == NULL) {
TRAP_ERROR();
return (EINVAL);
}
/* Make a node and try attach it to the type */
NG_ALLOC_NODE(node);
if (node == NULL) {
TRAP_ERROR();
return (ENOMEM);
}
node->nd_type = type;
NG_NODE_REF(node); /* note reference */
type->refs++;
mtx_init(&node->nd_input_queue.q_mtx, "ng_node", NULL, MTX_SPIN);
node->nd_input_queue.queue = NULL;
node->nd_input_queue.last = &node->nd_input_queue.queue;
node->nd_input_queue.q_flags = 0;
node->nd_input_queue.q_node = node;
/* Initialize hook list for new node */
LIST_INIT(&node->nd_hooks);
/* Link us into the node linked list */
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx);
LIST_INSERT_HEAD(&ng_nodelist, node, nd_nodes);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
/* get an ID and put us in the hash chain */
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_idhash_mtx);
for (;;) { /* wrap protection, even if silly */
node_p node2 = NULL;
node->nd_ID = nextID++; /* 137/second for 1 year before wrap */
/* Is there a problem with the new number? */
NG_IDHASH_FIND(node->nd_ID, node2); /* already taken? */
if ((node->nd_ID != 0) && (node2 == NULL)) {
break;
}
}
LIST_INSERT_HEAD(&ng_ID_hash[NG_IDHASH_FN(node->nd_ID)],
node, nd_idnodes);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_idhash_mtx);
/* Done */
*nodepp = node;
return (0);
}
/*
* Forceably start the shutdown process on a node. Either call
* it's shutdown method, or do the default shutdown if there is
* no type-specific method.
*
* We can only be called form a shutdown message, so we know we have
2001-01-11 22:22:52 +00:00
* a writer lock, and therefore exclusive access. It also means
* that we should not be on the work queue, but we check anyhow.
*
* Persistent node types must have a type-specific method which
2001-01-11 22:22:52 +00:00
* Allocates a new node in which case, this one is irretrievably going away,
* or cleans up anything it needs, and just makes the node valid again,
* in which case we allow the node to survive.
*
* XXX We need to think of how to tell a persistant node that we
* REALLY need to go away because the hardware has gone or we
* are rebooting.... etc.
*/
void
ng_rmnode(node_p node, hook_p dummy1, void *dummy2, int dummy3)
{
2001-01-11 22:22:52 +00:00
hook_p hook;
/* Check if it's already shutting down */
if ((node->nd_flags & NGF_CLOSING) != 0)
return;
if (node == &ng_deadnode) {
printf ("shutdown called on deadnode\n");
return;
}
/* Add an extra reference so it doesn't go away during this */
NG_NODE_REF(node);
/*
* Mark it invalid so any newcomers know not to try use it
* Also add our own mark so we can't recurse
* note that NGF_INVALID does not do this as it's also set during
* creation
*/
node->nd_flags |= NGF_INVALID|NGF_CLOSING;
/* If node has its pre-shutdown method, then call it first*/
if (node->nd_type && node->nd_type->close)
(*node->nd_type->close)(node);
2001-01-11 22:22:52 +00:00
/* Notify all remaining connected nodes to disconnect */
while ((hook = LIST_FIRST(&node->nd_hooks)) != NULL)
ng_destroy_hook(hook);
/*
* Drain the input queue forceably.
* it has no hooks so what's it going to do, bleed on someone?
* Theoretically we came here from a queue entry that was added
* Just before the queue was closed, so it should be empty anyway.
* Also removes us from worklist if needed.
*/
ng_flush_input_queue(&node->nd_input_queue);
/* Ask the type if it has anything to do in this case */
if (node->nd_type && node->nd_type->shutdown) {
(*node->nd_type->shutdown)(node);
if (NG_NODE_IS_VALID(node)) {
/*
* Well, blow me down if the node code hasn't declared
* that it doesn't want to die.
* Presumably it is a persistant node.
* If we REALLY want it to go away,
* e.g. hardware going away,
* Our caller should set NGF_REALLY_DIE in nd_flags.
*/
node->nd_flags &= ~(NGF_INVALID|NGF_CLOSING);
NG_NODE_UNREF(node); /* Assume they still have theirs */
return;
}
} else { /* do the default thing */
NG_NODE_UNREF(node);
}
ng_unname(node); /* basically a NOP these days */
/*
* Remove extra reference, possibly the last
* Possible other holders of references may include
* timeout callouts, but theoretically the node's supposed to
* have cancelled them. Possibly hardware dependencies may
* force a driver to 'linger' with a reference.
*/
NG_NODE_UNREF(node);
}
#ifdef NETGRAPH_DEBUG
void
ng_ref_node(node_p node)
{
_NG_NODE_REF(node);
}
#endif
/*
* Remove a reference to the node, possibly the last.
* deadnode always acts as it it were the last.
*/
int
ng_unref_node(node_p node)
{
int v;
if (node == &ng_deadnode) {
return (0);
}
do {
v = node->nd_refs - 1;
} while (! atomic_cmpset_int(&node->nd_refs, v + 1, v));
if (v == 0) { /* we were the last */
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx);
node->nd_type->refs--; /* XXX maybe should get types lock? */
LIST_REMOVE(node, nd_nodes);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_idhash_mtx);
LIST_REMOVE(node, nd_idnodes);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_idhash_mtx);
mtx_destroy(&node->nd_input_queue.q_mtx);
NG_FREE_NODE(node);
}
return (v);
}
/************************************************************************
Node ID handling
************************************************************************/
static node_p
ng_ID2noderef(ng_ID_t ID)
{
node_p node;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_idhash_mtx);
NG_IDHASH_FIND(ID, node);
if(node)
NG_NODE_REF(node);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_idhash_mtx);
return(node);
}
ng_ID_t
ng_node2ID(node_p node)
{
return (node ? NG_NODE_ID(node) : 0);
}
/************************************************************************
Node name handling
************************************************************************/
/*
* Assign a node a name. Once assigned, the name cannot be changed.
*/
int
ng_name_node(node_p node, const char *name)
{
int i;
node_p node2;
/* Check the name is valid */
for (i = 0; i < NG_NODESIZ; i++) {
if (name[i] == '\0' || name[i] == '.' || name[i] == ':')
break;
}
if (i == 0 || name[i] != '\0') {
TRAP_ERROR();
return (EINVAL);
}
if (ng_decodeidname(name) != 0) { /* valid IDs not allowed here */
TRAP_ERROR();
return (EINVAL);
}
/* Check the name isn't already being used */
if ((node2 = ng_name2noderef(node, name)) != NULL) {
NG_NODE_UNREF(node2);
TRAP_ERROR();
return (EADDRINUSE);
}
/* copy it */
strlcpy(NG_NODE_NAME(node), name, NG_NODESIZ);
return (0);
}
/*
* Find a node by absolute name. The name should NOT end with ':'
* The name "." means "this node" and "[xxx]" means "the node
* with ID (ie, at address) xxx".
*
* Returns the node if found, else NULL.
* Eventually should add something faster than a sequential search.
* Note it aquires a reference on the node so you can be sure it's still there.
*/
node_p
ng_name2noderef(node_p here, const char *name)
{
node_p node;
ng_ID_t temp;
/* "." means "this node" */
if (strcmp(name, ".") == 0) {
NG_NODE_REF(here);
return(here);
}
/* Check for name-by-ID */
if ((temp = ng_decodeidname(name)) != 0) {
return (ng_ID2noderef(temp));
}
/* Find node by name */
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx);
LIST_FOREACH(node, &ng_nodelist, nd_nodes) {
if (NG_NODE_IS_VALID(node)
&& NG_NODE_HAS_NAME(node)
&& (strcmp(NG_NODE_NAME(node), name) == 0)) {
break;
}
}
if (node)
NG_NODE_REF(node);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
return (node);
}
/*
* Decode an ID name, eg. "[f03034de]". Returns 0 if the
* string is not valid, otherwise returns the value.
*/
static ng_ID_t
ng_decodeidname(const char *name)
{
const int len = strlen(name);
char *eptr;
u_long val;
/* Check for proper length, brackets, no leading junk */
if ((len < 3)
|| (name[0] != '[')
|| (name[len - 1] != ']')
|| (!isxdigit(name[1]))) {
return ((ng_ID_t)0);
}
/* Decode number */
val = strtoul(name + 1, &eptr, 16);
if ((eptr - name != len - 1)
|| (val == ULONG_MAX)
|| (val == 0)) {
return ((ng_ID_t)0);
}
return (ng_ID_t)val;
}
/*
* Remove a name from a node. This should only be called
* when shutting down and removing the node.
* IF we allow name changing this may be more resurected.
*/
void
ng_unname(node_p node)
{
}
/************************************************************************
Hook routines
Names are not optional. Hooks are always connected, except for a
2001-01-11 22:22:52 +00:00
brief moment within these routines. On invalidation or during creation
they are connected to the 'dead' hook.
************************************************************************/
/*
* Remove a hook reference
*/
void
ng_unref_hook(hook_p hook)
{
int v;
if (hook == &ng_deadhook) {
return;
}
do {
v = hook->hk_refs;
} while (! atomic_cmpset_int(&hook->hk_refs, v, v - 1));
if (v == 1) { /* we were the last */
if (_NG_HOOK_NODE(hook)) { /* it'll probably be ng_deadnode */
_NG_NODE_UNREF((_NG_HOOK_NODE(hook)));
hook->hk_node = NULL;
}
NG_FREE_HOOK(hook);
}
}
/*
* Add an unconnected hook to a node. Only used internally.
2001-01-11 22:22:52 +00:00
* Assumes node is locked. (XXX not yet true )
*/
static int
ng_add_hook(node_p node, const char *name, hook_p *hookp)
{
hook_p hook;
int error = 0;
/* Check that the given name is good */
if (name == NULL) {
TRAP_ERROR();
return (EINVAL);
}
if (ng_findhook(node, name) != NULL) {
TRAP_ERROR();
return (EEXIST);
}
/* Allocate the hook and link it up */
NG_ALLOC_HOOK(hook);
if (hook == NULL) {
TRAP_ERROR();
return (ENOMEM);
}
2001-01-11 22:22:52 +00:00
hook->hk_refs = 1; /* add a reference for us to return */
hook->hk_flags = HK_INVALID;
2001-01-11 22:22:52 +00:00
hook->hk_peer = &ng_deadhook; /* start off this way */
hook->hk_node = node;
NG_NODE_REF(node); /* each hook counts as a reference */
2001-01-11 22:22:52 +00:00
/* Set hook name */
strlcpy(NG_HOOK_NAME(hook), name, NG_HOOKSIZ);
2001-01-11 22:22:52 +00:00
/*
* Check if the node type code has something to say about it
* If it fails, the unref of the hook will also unref the node.
*/
if (node->nd_type->newhook != NULL) {
if ((error = (*node->nd_type->newhook)(node, hook, name))) {
NG_HOOK_UNREF(hook); /* this frees the hook */
return (error);
}
}
/*
* The 'type' agrees so far, so go ahead and link it in.
* We'll ask again later when we actually connect the hooks.
*/
LIST_INSERT_HEAD(&node->nd_hooks, hook, hk_hooks);
node->nd_numhooks++;
2001-01-11 22:22:52 +00:00
NG_HOOK_REF(hook); /* one for the node */
if (hookp)
*hookp = hook;
2001-01-11 22:22:52 +00:00
return (0);
}
/*
* Find a hook
*
* Node types may supply their own optimized routines for finding
* hooks. If none is supplied, we just do a linear search.
2001-01-11 22:22:52 +00:00
* XXX Possibly we should add a reference to the hook?
*/
hook_p
ng_findhook(node_p node, const char *name)
{
hook_p hook;
if (node->nd_type->findhook != NULL)
return (*node->nd_type->findhook)(node, name);
LIST_FOREACH(hook, &node->nd_hooks, hk_hooks) {
if (NG_HOOK_IS_VALID(hook)
2001-01-11 04:13:46 +00:00
&& (strcmp(NG_HOOK_NAME(hook), name) == 0))
return (hook);
}
return (NULL);
}
/*
* Destroy a hook
*
* As hooks are always attached, this really destroys two hooks.
* The one given, and the one attached to it. Disconnect the hooks
2001-01-11 22:22:52 +00:00
* from each other first. We reconnect the peer hook to the 'dead'
* hook so that it can still exist after we depart. We then
* send the peer its own destroy message. This ensures that we only
* interact with the peer's structures when it is locked processing that
* message. We hold a reference to the peer hook so we are guaranteed that
* the peer hook and node are still going to exist until
* we are finished there as the hook holds a ref on the node.
* We run this same code again on the peer hook, but that time it is already
* attached to the 'dead' hook.
*
* This routine is called at all stages of hook creation
* on error detection and must be able to handle any such stage.
*/
void
ng_destroy_hook(hook_p hook)
{
hook_p peer = NG_HOOK_PEER(hook);
node_p node = NG_HOOK_NODE(hook);
if (hook == &ng_deadhook) { /* better safe than sorry */
printf("ng_destroy_hook called on deadhook\n");
return;
}
hook->hk_flags |= HK_INVALID; /* as soon as possible */
2001-01-11 22:22:52 +00:00
if (peer && (peer != &ng_deadhook)) {
/*
* Set the peer to point to ng_deadhook
* from this moment on we are effectively independent it.
* send it an rmhook message of it's own.
*/
peer->hk_peer = &ng_deadhook; /* They no longer know us */
hook->hk_peer = &ng_deadhook; /* Nor us, them */
if (NG_HOOK_NODE(peer) == &ng_deadnode) {
/*
* If it's already divorced from a node,
* just free it.
*/
/* nothing */
} else {
ng_rmhook_self(peer); /* Send it a surprise */
}
NG_HOOK_UNREF(peer); /* account for peer link */
NG_HOOK_UNREF(hook); /* account for peer link */
}
/*
* Remove the hook from the node's list to avoid possible recursion
* in case the disconnection results in node shutdown.
*/
if (node == &ng_deadnode) { /* happens if called from ng_con_nodes() */
return;
}
LIST_REMOVE(hook, hk_hooks);
node->nd_numhooks--;
if (node->nd_type->disconnect) {
/*
* The type handler may elect to destroy the node so don't
* trust its existance after this point. (except
* that we still hold a reference on it. (which we
* inherrited from the hook we are destroying)
*/
(*node->nd_type->disconnect) (hook);
}
/*
* Note that because we will point to ng_deadnode, the original node
* is not decremented automatically so we do that manually.
*/
_NG_HOOK_NODE(hook) = &ng_deadnode;
NG_NODE_UNREF(node); /* We no longer point to it so adjust count */
NG_HOOK_UNREF(hook); /* Account for linkage (in list) to node */
}
/*
* Take two hooks on a node and merge the connection so that the given node
* is effectively bypassed.
*/
int
ng_bypass(hook_p hook1, hook_p hook2)
{
if (hook1->hk_node != hook2->hk_node) {
TRAP_ERROR();
return (EINVAL);
}
hook1->hk_peer->hk_peer = hook2->hk_peer;
hook2->hk_peer->hk_peer = hook1->hk_peer;
2001-01-11 22:22:52 +00:00
hook1->hk_peer = &ng_deadhook;
hook2->hk_peer = &ng_deadhook;
/* XXX If we ever cache methods on hooks update them as well */
ng_destroy_hook(hook1);
ng_destroy_hook(hook2);
return (0);
}
/*
* Install a new netgraph type
*/
int
ng_newtype(struct ng_type *tp)
{
const size_t namelen = strlen(tp->name);
/* Check version and type name fields */
if ((tp->version != NG_ABI_VERSION)
|| (namelen == 0)
|| (namelen >= NG_TYPESIZ)) {
TRAP_ERROR();
if (tp->version != NG_ABI_VERSION) {
printf("Netgraph: Node type rejected. ABI mismatch. Suggest recompile\n");
}
return (EINVAL);
}
/* Check for name collision */
if (ng_findtype(tp->name) != NULL) {
TRAP_ERROR();
return (EEXIST);
}
/* Link in new type */
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_typelist_mtx);
LIST_INSERT_HEAD(&ng_typelist, tp, types);
tp->refs = 1; /* first ref is linked list */
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_typelist_mtx);
return (0);
}
/*
* unlink a netgraph type
* If no examples exist
*/
int
ng_rmtype(struct ng_type *tp)
{
/* Check for name collision */
if (tp->refs != 1) {
TRAP_ERROR();
return (EBUSY);
}
/* Unlink type */
mtx_lock(&ng_typelist_mtx);
LIST_REMOVE(tp, types);
mtx_unlock(&ng_typelist_mtx);
return (0);
}
/*
* Look for a type of the name given
*/
struct ng_type *
ng_findtype(const char *typename)
{
struct ng_type *type;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_typelist_mtx);
LIST_FOREACH(type, &ng_typelist, types) {
if (strcmp(type->name, typename) == 0)
break;
}
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_typelist_mtx);
return (type);
}
/************************************************************************
Composite routines
************************************************************************/
/*
* Connect two nodes using the specified hooks, using queued functions.
*/
static void
ng_con_part3(node_p node, hook_p hook, void *arg1, int arg2)
{
/*
* When we run, we know that the node 'node' is locked for us.
* Our caller has a reference on the hook.
* Our caller has a reference on the node.
* (In this case our caller is ng_apply_item() ).
* The peer hook has a reference on the hook.
* We are all set up except for the final call to the node, and
* the clearing of the INVALID flag.
*/
if (NG_HOOK_NODE(hook) == &ng_deadnode) {
/*
* The node must have been freed again since we last visited
* here. ng_destry_hook() has this effect but nothing else does.
* We should just release our references and
* free anything we can think of.
* Since we know it's been destroyed, and it's our caller
* that holds the references, just return.
*/
return ;
}
if (hook->hk_node->nd_type->connect) {
if ((*hook->hk_node->nd_type->connect) (hook)) {
ng_destroy_hook(hook); /* also zaps peer */
printf("failed in ng_con_part3()\n");
return ;
}
}
/*
* XXX this is wrong for SMP. Possibly we need
* to separate out 'create' and 'invalid' flags.
* should only set flags on hooks we have locked under our node.
*/
hook->hk_flags &= ~HK_INVALID;
return ;
}
static void
ng_con_part2(node_p node, hook_p hook, void *arg1, int arg2)
{
/*
* When we run, we know that the node 'node' is locked for us.
* Our caller has a reference on the hook.
* Our caller has a reference on the node.
* (In this case our caller is ng_apply_item() ).
* The peer hook has a reference on the hook.
* our node pointer points to the 'dead' node.
* First check the hook name is unique.
* Should not happen because we checked before queueing this.
*/
if (ng_findhook(node, NG_HOOK_NAME(hook)) != NULL) {
TRAP_ERROR();
ng_destroy_hook(hook); /* should destroy peer too */
printf("failed in ng_con_part2()\n");
return ;
}
2001-01-11 22:22:52 +00:00
/*
* Check if the node type code has something to say about it
* If it fails, the unref of the hook will also unref the attached node,
* however since that node is 'ng_deadnode' this will do nothing.
* The peer hook will also be destroyed.
2001-01-11 22:22:52 +00:00
*/
if (node->nd_type->newhook != NULL) {
if ((*node->nd_type->newhook)(node, hook, hook->hk_name)) {
ng_destroy_hook(hook); /* should destroy peer too */
printf("failed in ng_con_part2()\n");
return ;
}
}
/*
* The 'type' agrees so far, so go ahead and link it in.
* We'll ask again later when we actually connect the hooks.
*/
hook->hk_node = node; /* just overwrite ng_deadnode */
NG_NODE_REF(node); /* each hook counts as a reference */
LIST_INSERT_HEAD(&node->nd_hooks, hook, hk_hooks);
node->nd_numhooks++;
NG_HOOK_REF(hook); /* one for the node */
/*
* We now have a symetrical situation, where both hooks have been
* linked to their nodes, the newhook methods have been called
* And the references are all correct. The hooks are still marked
* as invalid, as we have not called the 'connect' methods
* yet.
* We can call the local one immediatly as we have the
* node locked, but we need to queue the remote one.
*/
if (hook->hk_node->nd_type->connect) {
if ((*hook->hk_node->nd_type->connect) (hook)) {
ng_destroy_hook(hook); /* also zaps peer */
printf("failed in ng_con_part2(A)\n");
return ;
}
}
if (ng_send_fn(hook->hk_peer->hk_node, hook->hk_peer,
&ng_con_part3, arg1, arg2)) {
printf("failed in ng_con_part2(B)");
ng_destroy_hook(hook); /* also zaps peer */
return ;
}
hook->hk_flags &= ~HK_INVALID; /* need both to be able to work */
return ;
}
/*
* Connect this node with another node. We assume that this node is
* currently locked, as we are only called from an NGM_CONNECT message.
*/
static int
ng_con_nodes(node_p node, const char *name, node_p node2, const char *name2)
{
int error;
hook_p hook;
hook_p hook2;
if (ng_findhook(node2, name2) != NULL) {
return(EEXIST);
}
2001-01-11 22:22:52 +00:00
if ((error = ng_add_hook(node, name, &hook))) /* gives us a ref */
return (error);
/* Allocate the other hook and link it up */
NG_ALLOC_HOOK(hook2);
if (hook2 == NULL) {
TRAP_ERROR();
ng_destroy_hook(hook); /* XXX check ref counts so far */
NG_HOOK_UNREF(hook); /* including our ref */
return (ENOMEM);
}
hook2->hk_refs = 1; /* start with a reference for us. */
hook2->hk_flags = HK_INVALID;
hook2->hk_peer = hook; /* Link the two together */
hook->hk_peer = hook2;
NG_HOOK_REF(hook); /* Add a ref for the peer to each*/
NG_HOOK_REF(hook2);
hook2->hk_node = &ng_deadnode;
strlcpy(NG_HOOK_NAME(hook2), name2, NG_HOOKSIZ);
/*
* Queue the function above.
* Procesing continues in that function in the lock context of
* the other node.
*/
ng_send_fn(node2, hook2, &ng_con_part2, NULL, 0);
NG_HOOK_UNREF(hook); /* Let each hook go if it wants to */
NG_HOOK_UNREF(hook2);
return (0);
}
/*
* Make a peer and connect.
* We assume that the local node is locked.
* The new node probably doesn't need a lock until
* it has a hook, because it cannot really have any work until then,
* but we should think about it a bit more.
*
* The problem may come if the other node also fires up
* some hardware or a timer or some other source of activation,
* also it may already get a command msg via it's ID.
*
* We could use the same method as ng_con_nodes() but we'd have
* to add ability to remove the node when failing. (Not hard, just
* make arg1 point to the node to remove).
* Unless of course we just ignore failure to connect and leave
* an unconnected node?
*/
static int
ng_mkpeer(node_p node, const char *name, const char *name2, char *type)
{
node_p node2;
hook_p hook1;
hook_p hook2;
int error;
if ((error = ng_make_node(type, &node2))) {
return (error);
}
if ((error = ng_add_hook(node, name, &hook1))) { /* gives us a ref */
ng_rmnode(node2, NULL, NULL, 0);
return (error);
}
if ((error = ng_add_hook(node2, name2, &hook2))) {
ng_rmnode(node2, NULL, NULL, 0);
ng_destroy_hook(hook1);
NG_HOOK_UNREF(hook1);
return (error);
}
/*
* Actually link the two hooks together.
*/
hook1->hk_peer = hook2;
hook2->hk_peer = hook1;
/* Each hook is referenced by the other */
NG_HOOK_REF(hook1);
NG_HOOK_REF(hook2);
/* Give each node the opportunity to veto the pending connection */
if (hook1->hk_node->nd_type->connect) {
error = (*hook1->hk_node->nd_type->connect) (hook1);
}
if ((error == 0) && hook2->hk_node->nd_type->connect) {
error = (*hook2->hk_node->nd_type->connect) (hook2);
}
2001-01-11 22:22:52 +00:00
/*
* drop the references we were holding on the two hooks.
*/
if (error) {
ng_destroy_hook(hook2); /* also zaps hook1 */
ng_rmnode(node2, NULL, NULL, 0);
} else {
/* As a last act, allow the hooks to be used */
hook1->hk_flags &= ~HK_INVALID;
hook2->hk_flags &= ~HK_INVALID;
}
NG_HOOK_UNREF(hook1);
2001-01-11 22:22:52 +00:00
NG_HOOK_UNREF(hook2);
return (error);
}
/************************************************************************
Utility routines to send self messages
************************************************************************/
/* Shut this node down as soon as everyone is clear of it */
/* Should add arg "immediatly" to jump the queue */
int
ng_rmnode_self(node_p node)
{
int error;
if (node == &ng_deadnode)
return (0);
node->nd_flags |= NGF_INVALID;
if (node->nd_flags & NGF_CLOSING)
return (0);
error = ng_send_fn(node, NULL, &ng_rmnode, NULL, 0);
return (error);
}
static void
ng_rmhook_part2(node_p node, hook_p hook, void *arg1, int arg2)
{
ng_destroy_hook(hook);
return ;
}
int
ng_rmhook_self(hook_p hook)
{
int error;
node_p node = NG_HOOK_NODE(hook);
if (node == &ng_deadnode)
return (0);
error = ng_send_fn(node, hook, &ng_rmhook_part2, NULL, 0);
return (error);
}
/***********************************************************************
* Parse and verify a string of the form: <NODE:><PATH>
*
* Such a string can refer to a specific node or a specific hook
* on a specific node, depending on how you look at it. In the
* latter case, the PATH component must not end in a dot.
*
* Both <NODE:> and <PATH> are optional. The <PATH> is a string
* of hook names separated by dots. This breaks out the original
* string, setting *nodep to "NODE" (or NULL if none) and *pathp
* to "PATH" (or NULL if degenerate). Also, *hookp will point to
* the final hook component of <PATH>, if any, otherwise NULL.
*
* This returns -1 if the path is malformed. The char ** are optional.
***********************************************************************/
int
ng_path_parse(char *addr, char **nodep, char **pathp, char **hookp)
{
char *node, *path, *hook;
int k;
/*
* Extract absolute NODE, if any
*/
for (path = addr; *path && *path != ':'; path++);
if (*path) {
node = addr; /* Here's the NODE */
*path++ = '\0'; /* Here's the PATH */
/* Node name must not be empty */
if (!*node)
return -1;
/* A name of "." is OK; otherwise '.' not allowed */
if (strcmp(node, ".") != 0) {
for (k = 0; node[k]; k++)
if (node[k] == '.')
return -1;
}
} else {
node = NULL; /* No absolute NODE */
path = addr; /* Here's the PATH */
}
/* Snoop for illegal characters in PATH */
for (k = 0; path[k]; k++)
if (path[k] == ':')
return -1;
/* Check for no repeated dots in PATH */
for (k = 0; path[k]; k++)
if (path[k] == '.' && path[k + 1] == '.')
return -1;
/* Remove extra (degenerate) dots from beginning or end of PATH */
if (path[0] == '.')
path++;
if (*path && path[strlen(path) - 1] == '.')
path[strlen(path) - 1] = 0;
/* If PATH has a dot, then we're not talking about a hook */
if (*path) {
for (hook = path, k = 0; path[k]; k++)
if (path[k] == '.') {
hook = NULL;
break;
}
} else
path = hook = NULL;
/* Done */
if (nodep)
*nodep = node;
if (pathp)
*pathp = path;
if (hookp)
*hookp = hook;
return (0);
}
/*
* Given a path, which may be absolute or relative, and a starting node,
* return the destination node.
*/
int
ng_path2noderef(node_p here, const char *address,
node_p *destp, hook_p *lasthook)
{
char fullpath[NG_PATHSIZ];
char *nodename, *path, pbuf[2];
node_p node, oldnode;
char *cp;
hook_p hook = NULL;
/* Initialize */
if (destp == NULL) {
TRAP_ERROR();
return EINVAL;
}
*destp = NULL;
/* Make a writable copy of address for ng_path_parse() */
strncpy(fullpath, address, sizeof(fullpath) - 1);
fullpath[sizeof(fullpath) - 1] = '\0';
/* Parse out node and sequence of hooks */
if (ng_path_parse(fullpath, &nodename, &path, NULL) < 0) {
TRAP_ERROR();
return EINVAL;
}
if (path == NULL) {
pbuf[0] = '.'; /* Needs to be writable */
pbuf[1] = '\0';
path = pbuf;
}
/*
* For an absolute address, jump to the starting node.
* Note that this holds a reference on the node for us.
* Don't forget to drop the reference if we don't need it.
*/
if (nodename) {
node = ng_name2noderef(here, nodename);
if (node == NULL) {
TRAP_ERROR();
return (ENOENT);
}
} else {
if (here == NULL) {
TRAP_ERROR();
return (EINVAL);
}
node = here;
NG_NODE_REF(node);
}
/*
* Now follow the sequence of hooks
* XXX
* We actually cannot guarantee that the sequence
* is not being demolished as we crawl along it
* without extra-ordinary locking etc.
* So this is a bit dodgy to say the least.
* We can probably hold up some things by holding
* the nodelist mutex for the time of this
* crawl if we wanted.. At least that way we wouldn't have to
* worry about the nodes dissappearing, but the hooks would still
* be a problem.
*/
for (cp = path; node != NULL && *cp != '\0'; ) {
char *segment;
/*
* Break out the next path segment. Replace the dot we just
* found with a NUL; "cp" points to the next segment (or the
* NUL at the end).
*/
for (segment = cp; *cp != '\0'; cp++) {
if (*cp == '.') {
*cp++ = '\0';
break;
}
}
/* Empty segment */
if (*segment == '\0')
continue;
/* We have a segment, so look for a hook by that name */
hook = ng_findhook(node, segment);
/* Can't get there from here... */
if (hook == NULL
|| NG_HOOK_PEER(hook) == NULL
|| NG_HOOK_NOT_VALID(hook)
|| NG_HOOK_NOT_VALID(NG_HOOK_PEER(hook))) {
TRAP_ERROR();
NG_NODE_UNREF(node);
#if 0
printf("hooknotvalid %s %s %d %d %d %d ",
path,
segment,
hook == NULL,
NG_HOOK_PEER(hook) == NULL,
NG_HOOK_NOT_VALID(hook),
NG_HOOK_NOT_VALID(NG_HOOK_PEER(hook)));
#endif
return (ENOENT);
}
/*
* Hop on over to the next node
* XXX
* Big race conditions here as hooks and nodes go away
* *** Idea.. store an ng_ID_t in each hook and use that
* instead of the direct hook in this crawl?
*/
oldnode = node;
if ((node = NG_PEER_NODE(hook)))
NG_NODE_REF(node); /* XXX RACE */
NG_NODE_UNREF(oldnode); /* XXX another race */
if (NG_NODE_NOT_VALID(node)) {
NG_NODE_UNREF(node); /* XXX more races */
node = NULL;
}
}
/* If node somehow missing, fail here (probably this is not needed) */
if (node == NULL) {
TRAP_ERROR();
return (ENXIO);
}
/* Done */
*destp = node;
if (lasthook != NULL)
*lasthook = (hook ? NG_HOOK_PEER(hook) : NULL);
return (0);
}
/***************************************************************\
* Input queue handling.
* All activities are submitted to the node via the input queue
* which implements a multiple-reader/single-writer gate.
* Items which cannot be handled immeditly are queued.
*
* read-write queue locking inline functions *
\***************************************************************/
static __inline item_p ng_dequeue(struct ng_queue * ngq);
static __inline item_p ng_acquire_read(struct ng_queue * ngq,
item_p item);
static __inline item_p ng_acquire_write(struct ng_queue * ngq,
item_p item);
static __inline void ng_leave_read(struct ng_queue * ngq);
static __inline void ng_leave_write(struct ng_queue * ngq);
static __inline void ng_queue_rw(struct ng_queue * ngq,
item_p item, int rw);
/*
* Definition of the bits fields in the ng_queue flag word.
* Defined here rather than in netgraph.h because no-one should fiddle
* with them.
*
* The ordering here may be important! don't shuffle these.
*/
/*-
Safety Barrier--------+ (adjustable to suit taste) (not used yet)
|
V
+-------+-------+-------+-------+-------+-------+-------+-------+
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
| |A|c|t|i|v|e| |R|e|a|d|e|r| |C|o|u|n|t| | | | | | | | | |R|A|W|
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | |P|W|P|
+-------+-------+-------+-------+-------+-------+-------+-------+
\___________________________ ____________________________/ | | |
V | | |
[active reader count] | | |
| | |
Read Pending ------------------------------------+ | |
| |
Active Writer -------------------------------------+ |
|
Write Pending ---------------------------------------+
*/
#define WRITE_PENDING 0x00000001
#define WRITER_ACTIVE 0x00000002
#define READ_PENDING 0x00000004
#define READER_INCREMENT 0x00000008
#define READER_MASK 0xfffffff0 /* Not valid if WRITER_ACTIVE is set */
#define SAFETY_BARRIER 0x00100000 /* 64K items queued should be enough */
/* Defines of more elaborate states on the queue */
/* Mask of bits a read cares about */
#define NGQ_RMASK (WRITE_PENDING|WRITER_ACTIVE|READ_PENDING)
/* Mask of bits a write cares about */
#define NGQ_WMASK (NGQ_RMASK|READER_MASK)
/* tests to decide if we could get a read or write off the queue */
#define CAN_GET_READ(flag) ((flag & NGQ_RMASK) == READ_PENDING)
#define CAN_GET_WRITE(flag) ((flag & NGQ_WMASK) == WRITE_PENDING)
/* Is there a chance of getting ANY work off the queue? */
#define CAN_GET_WORK(flag) (CAN_GET_READ(flag) || CAN_GET_WRITE(flag))
/*
* Taking into account the current state of the queue and node, possibly take
* the next entry off the queue and return it. Return NULL if there was
* nothing we could return, either because there really was nothing there, or
* because the node was in a state where it cannot yet process the next item
* on the queue.
*
* This MUST MUST MUST be called with the mutex held.
*/
static __inline item_p
ng_dequeue(struct ng_queue *ngq)
{
item_p item;
u_int add_arg;
mtx_assert(&ngq->q_mtx, MA_OWNED);
if (CAN_GET_READ(ngq->q_flags)) {
/*
* Head of queue is a reader and we have no write active.
* We don't care how many readers are already active.
* Adjust the flags for the item we are about to dequeue.
* Add the correct increment for the reader count as well.
*/
add_arg = (READER_INCREMENT - READ_PENDING);
} else if (CAN_GET_WRITE(ngq->q_flags)) {
/*
* There is a pending write, no readers and no active writer.
* This means we can go ahead with the pending writer. Note
* the fact that we now have a writer, ready for when we take
* it off the queue.
*
* We don't need to worry about a possible collision with the
* fasttrack reader.
*
* The fasttrack thread may take a long time to discover that we
* are running so we would have an inconsistent state in the
* flags for a while. Since we ignore the reader count
* entirely when the WRITER_ACTIVE flag is set, this should
* not matter (in fact it is defined that way). If it tests
* the flag before this operation, the WRITE_PENDING flag
* will make it fail, and if it tests it later, the
* WRITER_ACTIVE flag will do the same. If it is SO slow that
* we have actually completed the operation, and neither flag
* is set (nor the READ_PENDING) by the time that it tests
* the flags, then it is actually ok for it to continue. If
* it completes and we've finished and the read pending is
* set it still fails.
*
* So we can just ignore it, as long as we can ensure that the
* transition from WRITE_PENDING state to the WRITER_ACTIVE
* state is atomic.
*
* After failing, first it will be held back by the mutex, then
* when it can proceed, it will queue its request, then it
* would arrive at this function. Usually it will have to
* leave empty handed because the ACTIVE WRITER bit will be
* set.
*
* Adjust the flags for the item we are about to dequeue
* and for the new active writer.
*/
add_arg = (WRITER_ACTIVE - WRITE_PENDING);
/*
* We want to write "active writer, no readers " Now go make
* it true. In fact there may be a number in the readers
* count but we know it is not true and will be fixed soon.
* We will fix the flags for the next pending entry in a
* moment.
*/
} else {
/*
* We can't dequeue anything.. return and say so. Probably we
* have a write pending and the readers count is non zero. If
* we got here because a reader hit us just at the wrong
* moment with the fasttrack code, and put us in a strange
* state, then it will be through in just a moment, (as soon
* as we release the mutex) and keep things moving.
* Make sure we remove ourselves from the work queue.
*/
ng_worklist_remove(ngq->q_node);
return (0);
}
/*
* Now we dequeue the request (whatever it may be) and correct the
* pending flags and the next and last pointers.
*/
item = ngq->queue;
ngq->queue = item->el_next;
if (ngq->last == &(item->el_next)) {
/*
* that was the last entry in the queue so set the 'last
* pointer up correctly and make sure the pending flags are
* clear.
*/
ngq->last = &(ngq->queue);
/*
* Whatever flag was set will be cleared and
* the new acive field will be set by the add as well,
* so we don't need to change add_arg.
* But we know we don't need to be on the work list.
*/
atomic_add_long(&ngq->q_flags, add_arg);
ng_worklist_remove(ngq->q_node);
} else {
/*
* Since there is something on the queue, note what it is
* in the flags word.
*/
if ((ngq->queue->el_flags & NGQF_RW) == NGQF_READER) {
add_arg += READ_PENDING;
} else {
add_arg += WRITE_PENDING;
}
atomic_add_long(&ngq->q_flags, add_arg);
/*
* If we see more doable work, make sure we are
* on the work queue.
*/
if (CAN_GET_WORK(ngq->q_flags)) {
ng_setisr(ngq->q_node);
}
}
/*
* We have successfully cleared the old pending flag, set the new one
* if it is needed, and incremented the appropriate active field.
* (all in one atomic addition.. )
*/
return (item);
}
/*
* Queue a packet to be picked up by someone else.
* We really don't care who, but we can't or don't want to hang around
* to process it ourselves. We are probably an interrupt routine..
* 1 = writer, 0 = reader
*/
#define NGQRW_R 0
#define NGQRW_W 1
static __inline void
ng_queue_rw(struct ng_queue * ngq, item_p item, int rw)
{
mtx_assert(&ngq->q_mtx, MA_OWNED);
item->el_next = NULL; /* maybe not needed */
*ngq->last = item;
/*
* If it was the first item in the queue then we need to
* set the last pointer and the type flags.
*/
if (ngq->last == &(ngq->queue)) {
/*
* When called with constants for rw, the optimiser will
* remove the unneeded branch below.
*/
if (rw == NGQRW_W) {
atomic_add_long(&ngq->q_flags, WRITE_PENDING);
} else {
atomic_add_long(&ngq->q_flags, READ_PENDING);
}
}
ngq->last = &(item->el_next);
}
/*
* This function 'cheats' in that it first tries to 'grab' the use of the
* node, without going through the mutex. We can do this becasue of the
* semantics of the lock. The semantics include a clause that says that the
* value of the readers count is invalid if the WRITER_ACTIVE flag is set. It
* also says that the WRITER_ACTIVE flag cannot be set if the readers count
* is not zero. Note that this talks about what is valid to SET the
* WRITER_ACTIVE flag, because from the moment it is set, the value if the
* reader count is immaterial, and not valid. The two 'pending' flags have a
* similar effect, in that If they are orthogonal to the two active fields in
* how they are set, but if either is set, the attempted 'grab' need to be
* backed out because there is earlier work, and we maintain ordering in the
* queue. The result of this is that the reader request can try obtain use of
* the node with only a single atomic addition, and without any of the mutex
* overhead. If this fails the operation degenerates to the same as for other
* cases.
*
*/
static __inline item_p
ng_acquire_read(struct ng_queue *ngq, item_p item)
{
/* ######### Hack alert ######### */
atomic_add_long(&ngq->q_flags, READER_INCREMENT);
if ((ngq->q_flags & NGQ_RMASK) == 0) {
/* Successfully grabbed node */
return (item);
}
/* undo the damage if we didn't succeed */
atomic_subtract_long(&ngq->q_flags, READER_INCREMENT);
/* ######### End Hack alert ######### */
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock_spin((&ngq->q_mtx));
/*
* Try again. Another processor (or interrupt for that matter) may
* have removed the last queued item that was stopping us from
* running, between the previous test, and the moment that we took
* the mutex. (Or maybe a writer completed.)
*/
if ((ngq->q_flags & NGQ_RMASK) == 0) {
atomic_add_long(&ngq->q_flags, READER_INCREMENT);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin((&ngq->q_mtx));
return (item);
}
/*
* and queue the request for later.
*/
item->el_flags |= NGQF_READER;
ng_queue_rw(ngq, item, NGQRW_R);
Problem description: At the end of ng_snd_item(), node queue is processed. In certain netgraph setups deep recursive calls can occur. For example this happens, when two nodes are connected and can send items to each other in both directions. If, for some reason, both nodes have a lot of items in their queues, then the processing thread will recurse between these two nodes, delivering items left and right, going deeper in the stack. Other setups can suffer from deep recursion, too. The following factors can influence risk of deep netgraph call: - periodical write-access events on node - combination of slow link and fast one in one graph - net.inet.ip.fastforwarding Changes made: - In ng_acquire_{read,write}() do not dequeue another item. Instead, call ng_setisr() for this node. - At the end of ng_snd_item(), do not process queue. Call ng_setisr(), if there are any dequeueable items on node queue. - In ng_setisr() narrow worklist mutex holding. - In ng_setisr() assert queue mutex. Theoretically, the first two changes should negatively affect performance. To check this, some profiling was made: 1) In general real tasks, no noticable performance difference was found. 2) The following test was made: two multithreaded nodes and one single-threaded were connected into a ring. A large queues of packets were sent around this ring. Time to pass the ring N times was measured. This is a very vacuous test: no items/mbufs are allocated, no upcalls or downcalls outside of netgraph. It doesn't represent a real load, it is a stress test for ng_acquire_{read,write}() and item queueing functions. Surprisingly, the performance impact was positive! New code is 13% faster on UP and 17% faster on SMP, in this particular test. The problem was originally found, described, analyzed and original patch was written by Roselyn Lee from Vernier Networks. Thanks! Submitted by: Roselyn Lee <rosel verniernetworks com>
2005-07-21 12:08:37 +00:00
if (CAN_GET_WORK(ngq->q_flags))
ng_setisr(ngq->q_node);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin(&(ngq->q_mtx));
Problem description: At the end of ng_snd_item(), node queue is processed. In certain netgraph setups deep recursive calls can occur. For example this happens, when two nodes are connected and can send items to each other in both directions. If, for some reason, both nodes have a lot of items in their queues, then the processing thread will recurse between these two nodes, delivering items left and right, going deeper in the stack. Other setups can suffer from deep recursion, too. The following factors can influence risk of deep netgraph call: - periodical write-access events on node - combination of slow link and fast one in one graph - net.inet.ip.fastforwarding Changes made: - In ng_acquire_{read,write}() do not dequeue another item. Instead, call ng_setisr() for this node. - At the end of ng_snd_item(), do not process queue. Call ng_setisr(), if there are any dequeueable items on node queue. - In ng_setisr() narrow worklist mutex holding. - In ng_setisr() assert queue mutex. Theoretically, the first two changes should negatively affect performance. To check this, some profiling was made: 1) In general real tasks, no noticable performance difference was found. 2) The following test was made: two multithreaded nodes and one single-threaded were connected into a ring. A large queues of packets were sent around this ring. Time to pass the ring N times was measured. This is a very vacuous test: no items/mbufs are allocated, no upcalls or downcalls outside of netgraph. It doesn't represent a real load, it is a stress test for ng_acquire_{read,write}() and item queueing functions. Surprisingly, the performance impact was positive! New code is 13% faster on UP and 17% faster on SMP, in this particular test. The problem was originally found, described, analyzed and original patch was written by Roselyn Lee from Vernier Networks. Thanks! Submitted by: Roselyn Lee <rosel verniernetworks com>
2005-07-21 12:08:37 +00:00
return (NULL);
}
static __inline item_p
ng_acquire_write(struct ng_queue *ngq, item_p item)
{
restart:
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock_spin(&(ngq->q_mtx));
/*
* If there are no readers, no writer, and no pending packets, then
* we can just go ahead. In all other situations we need to queue the
* request
*/
if ((ngq->q_flags & NGQ_WMASK) == 0) {
atomic_add_long(&ngq->q_flags, WRITER_ACTIVE);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin((&ngq->q_mtx));
if (ngq->q_flags & READER_MASK) {
/* Collision with fast-track reader */
atomic_subtract_long(&ngq->q_flags, WRITER_ACTIVE);
goto restart;
}
return (item);
}
/*
* and queue the request for later.
*/
item->el_flags &= ~NGQF_RW;
ng_queue_rw(ngq, item, NGQRW_W);
Problem description: At the end of ng_snd_item(), node queue is processed. In certain netgraph setups deep recursive calls can occur. For example this happens, when two nodes are connected and can send items to each other in both directions. If, for some reason, both nodes have a lot of items in their queues, then the processing thread will recurse between these two nodes, delivering items left and right, going deeper in the stack. Other setups can suffer from deep recursion, too. The following factors can influence risk of deep netgraph call: - periodical write-access events on node - combination of slow link and fast one in one graph - net.inet.ip.fastforwarding Changes made: - In ng_acquire_{read,write}() do not dequeue another item. Instead, call ng_setisr() for this node. - At the end of ng_snd_item(), do not process queue. Call ng_setisr(), if there are any dequeueable items on node queue. - In ng_setisr() narrow worklist mutex holding. - In ng_setisr() assert queue mutex. Theoretically, the first two changes should negatively affect performance. To check this, some profiling was made: 1) In general real tasks, no noticable performance difference was found. 2) The following test was made: two multithreaded nodes and one single-threaded were connected into a ring. A large queues of packets were sent around this ring. Time to pass the ring N times was measured. This is a very vacuous test: no items/mbufs are allocated, no upcalls or downcalls outside of netgraph. It doesn't represent a real load, it is a stress test for ng_acquire_{read,write}() and item queueing functions. Surprisingly, the performance impact was positive! New code is 13% faster on UP and 17% faster on SMP, in this particular test. The problem was originally found, described, analyzed and original patch was written by Roselyn Lee from Vernier Networks. Thanks! Submitted by: Roselyn Lee <rosel verniernetworks com>
2005-07-21 12:08:37 +00:00
if (CAN_GET_WORK(ngq->q_flags))
ng_setisr(ngq->q_node);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin(&(ngq->q_mtx));
Problem description: At the end of ng_snd_item(), node queue is processed. In certain netgraph setups deep recursive calls can occur. For example this happens, when two nodes are connected and can send items to each other in both directions. If, for some reason, both nodes have a lot of items in their queues, then the processing thread will recurse between these two nodes, delivering items left and right, going deeper in the stack. Other setups can suffer from deep recursion, too. The following factors can influence risk of deep netgraph call: - periodical write-access events on node - combination of slow link and fast one in one graph - net.inet.ip.fastforwarding Changes made: - In ng_acquire_{read,write}() do not dequeue another item. Instead, call ng_setisr() for this node. - At the end of ng_snd_item(), do not process queue. Call ng_setisr(), if there are any dequeueable items on node queue. - In ng_setisr() narrow worklist mutex holding. - In ng_setisr() assert queue mutex. Theoretically, the first two changes should negatively affect performance. To check this, some profiling was made: 1) In general real tasks, no noticable performance difference was found. 2) The following test was made: two multithreaded nodes and one single-threaded were connected into a ring. A large queues of packets were sent around this ring. Time to pass the ring N times was measured. This is a very vacuous test: no items/mbufs are allocated, no upcalls or downcalls outside of netgraph. It doesn't represent a real load, it is a stress test for ng_acquire_{read,write}() and item queueing functions. Surprisingly, the performance impact was positive! New code is 13% faster on UP and 17% faster on SMP, in this particular test. The problem was originally found, described, analyzed and original patch was written by Roselyn Lee from Vernier Networks. Thanks! Submitted by: Roselyn Lee <rosel verniernetworks com>
2005-07-21 12:08:37 +00:00
return (NULL);
}
static __inline void
ng_leave_read(struct ng_queue *ngq)
{
atomic_subtract_long(&ngq->q_flags, READER_INCREMENT);
}
static __inline void
ng_leave_write(struct ng_queue *ngq)
{
atomic_subtract_long(&ngq->q_flags, WRITER_ACTIVE);
}
static void
ng_flush_input_queue(struct ng_queue * ngq)
{
item_p item;
u_int add_arg;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock_spin(&ngq->q_mtx);
for (;;) {
/* Now take a look at what's on the queue */
if (ngq->q_flags & READ_PENDING) {
add_arg = -READ_PENDING;
} else if (ngq->q_flags & WRITE_PENDING) {
add_arg = -WRITE_PENDING;
} else {
break;
}
item = ngq->queue;
ngq->queue = item->el_next;
if (ngq->last == &(item->el_next)) {
ngq->last = &(ngq->queue);
} else {
if ((ngq->queue->el_flags & NGQF_RW) == NGQF_READER) {
add_arg += READ_PENDING;
} else {
add_arg += WRITE_PENDING;
}
}
atomic_add_long(&ngq->q_flags, add_arg);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin(&ngq->q_mtx);
NG_FREE_ITEM(item);
mtx_lock_spin(&ngq->q_mtx);
}
/*
* Take us off the work queue if we are there.
* We definatly have no work to be done.
*/
ng_worklist_remove(ngq->q_node);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin(&ngq->q_mtx);
}
/***********************************************************************
* Externally visible method for sending or queueing messages or data.
***********************************************************************/
/*
* The module code should have filled out the item correctly by this stage:
* Common:
* reference to destination node.
* Reference to destination rcv hook if relevant.
* Data:
* pointer to mbuf
* Control_Message:
* pointer to msg.
* ID of original sender node. (return address)
* Function:
* Function pointer
* void * argument
* integer argument
*
* The nodes have several routines and macros to help with this task:
*/
int
ng_snd_item(item_p item, int flags)
{
hook_p hook = NGI_HOOK(item);
node_p node = NGI_NODE(item);
int queue, rw;
int error = 0, ierror;
item_p oitem;
struct ng_queue * ngq = &node->nd_input_queue;
#ifdef NETGRAPH_DEBUG
_ngi_check(item, __FILE__, __LINE__);
#endif
queue = (flags & NG_QUEUE) ? 1 : 0;
if (item == NULL) {
TRAP_ERROR();
return (EINVAL); /* failed to get queue element */
}
if (node == NULL) {
NG_FREE_ITEM(item);
TRAP_ERROR();
return (EINVAL); /* No address */
}
switch(item->el_flags & NGQF_TYPE) {
case NGQF_DATA:
/*
* DATA MESSAGE
* Delivered to a node via a non-optional hook.
* Both should be present in the item even though
* the node is derivable from the hook.
* References are held on both by the item.
*/
/* Protect nodes from sending NULL pointers
* to each other
*/
if (NGI_M(item) == NULL)
return (EINVAL);
CHECK_DATA_MBUF(NGI_M(item));
if (hook == NULL) {
NG_FREE_ITEM(item);
TRAP_ERROR();
return(EINVAL);
}
if ((NG_HOOK_NOT_VALID(hook))
|| (NG_NODE_NOT_VALID(NG_HOOK_NODE(hook)))) {
NG_FREE_ITEM(item);
return (ENOTCONN);
}
if ((hook->hk_flags & HK_QUEUE)) {
queue = 1;
}
/* By default data is a reader in the locking scheme */
item->el_flags |= NGQF_READER;
rw = NGQRW_R;
break;
case NGQF_MESG:
/*
* CONTROL MESSAGE
* Delivered to a node.
* Hook is optional.
* References are held by the item on the node and
* the hook if it is present.
*/
if (hook && (hook->hk_flags & HK_QUEUE)) {
queue = 1;
}
/* Data messages count as writers unles explicitly exempted */
if (NGI_MSG(item)->header.cmd & NGM_READONLY) {
item->el_flags |= NGQF_READER;
rw = NGQRW_R;
} else {
item->el_flags &= ~NGQF_RW;
rw = NGQRW_W;
}
break;
case NGQF_FN:
item->el_flags &= ~NGQF_RW;
rw = NGQRW_W;
break;
default:
NG_FREE_ITEM(item);
TRAP_ERROR();
return (EINVAL);
}
/*
* If the node specifies single threading, force writer semantics
* Similarly the node may say one hook always produces writers.
* These are over-rides.
*/
if ((node->nd_flags & NGF_FORCE_WRITER)
|| (hook && (hook->hk_flags & HK_FORCE_WRITER))) {
rw = NGQRW_W;
item->el_flags &= ~NGQF_READER;
}
if (queue) {
/* Put it on the queue for that node*/
#ifdef NETGRAPH_DEBUG
_ngi_check(item, __FILE__, __LINE__);
#endif
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock_spin(&(ngq->q_mtx));
ng_queue_rw(ngq, item, rw);
/*
* If there are active elements then we can rely on
* them. if not we should not rely on another packet
* coming here by another path,
* so it is best to put us in the netisr list.
* We can take the worklist lock with the node locked
* BUT NOT THE REVERSE!
*/
if (CAN_GET_WORK(ngq->q_flags)) {
ng_setisr(node);
}
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin(&(ngq->q_mtx));
if (flags & NG_PROGRESS)
return (EINPROGRESS);
else
return (0);
}
/*
* Take a queue item and a node and see if we can apply the item to
* the node. We may end up getting a different item to apply instead.
* Will allow for a piggyback reply only in the case where
* there is no queueing.
*/
oitem = item;
/*
* We already decided how we will be queueud or treated.
* Try get the appropriate operating permission.
*/
if (rw == NGQRW_R) {
item = ng_acquire_read(ngq, item);
} else {
item = ng_acquire_write(ngq, item);
}
/*
* May have come back with a different item.
* or maybe none at all. The one we started with will
* have been queued in thises cases.
*/
if (item == NULL) {
if (flags & NG_PROGRESS)
return (EINPROGRESS);
else
return (0);
}
#ifdef NETGRAPH_DEBUG
_ngi_check(item, __FILE__, __LINE__);
#endif
/*
* Take over the reference frm the item.
* Hold it until the called function returns.
*/
NGI_GET_NODE(item, node); /* zaps stored node */
ierror = ng_apply_item(node, item); /* drops r/w lock when done */
/* only return an error if it was our initial item.. (compat hack) */
if (oitem == item) {
error = ierror;
}
/*
* If the node goes away when we remove the reference,
2001-08-21 13:20:02 +00:00
* whatever we just did caused it.. whatever we do, DO NOT
* access the node again!
*/
if (NG_NODE_UNREF(node) == 0) {
return (error);
}
Problem description: At the end of ng_snd_item(), node queue is processed. In certain netgraph setups deep recursive calls can occur. For example this happens, when two nodes are connected and can send items to each other in both directions. If, for some reason, both nodes have a lot of items in their queues, then the processing thread will recurse between these two nodes, delivering items left and right, going deeper in the stack. Other setups can suffer from deep recursion, too. The following factors can influence risk of deep netgraph call: - periodical write-access events on node - combination of slow link and fast one in one graph - net.inet.ip.fastforwarding Changes made: - In ng_acquire_{read,write}() do not dequeue another item. Instead, call ng_setisr() for this node. - At the end of ng_snd_item(), do not process queue. Call ng_setisr(), if there are any dequeueable items on node queue. - In ng_setisr() narrow worklist mutex holding. - In ng_setisr() assert queue mutex. Theoretically, the first two changes should negatively affect performance. To check this, some profiling was made: 1) In general real tasks, no noticable performance difference was found. 2) The following test was made: two multithreaded nodes and one single-threaded were connected into a ring. A large queues of packets were sent around this ring. Time to pass the ring N times was measured. This is a very vacuous test: no items/mbufs are allocated, no upcalls or downcalls outside of netgraph. It doesn't represent a real load, it is a stress test for ng_acquire_{read,write}() and item queueing functions. Surprisingly, the performance impact was positive! New code is 13% faster on UP and 17% faster on SMP, in this particular test. The problem was originally found, described, analyzed and original patch was written by Roselyn Lee from Vernier Networks. Thanks! Submitted by: Roselyn Lee <rosel verniernetworks com>
2005-07-21 12:08:37 +00:00
mtx_lock_spin(&(ngq->q_mtx));
if (CAN_GET_WORK(ngq->q_flags))
ng_setisr(ngq->q_node);
mtx_unlock_spin(&(ngq->q_mtx));
return (error);
}
/*
* We have an item that was possibly queued somewhere.
* It should contain all the information needed
* to run it on the appropriate node/hook.
*/
static int
ng_apply_item(node_p node, item_p item)
{
hook_p hook;
int was_reader = ((item->el_flags & NGQF_RW));
int error = 0;
ng_rcvdata_t *rcvdata;
ng_rcvmsg_t *rcvmsg;
ng_apply_t *apply = NULL;
void *context = NULL;
NGI_GET_HOOK(item, hook); /* clears stored hook */
#ifdef NETGRAPH_DEBUG
_ngi_check(item, __FILE__, __LINE__);
#endif
/*
* If item has apply callback, store it. Clear callback
* immediately, two avoid another call in case if
* item would be reused by destination node.
*/
if (item->apply != NULL) {
apply = item->apply;
context = item->context;
item->apply = NULL;
}
switch (item->el_flags & NGQF_TYPE) {
case NGQF_DATA:
/*
* Check things are still ok as when we were queued.
*/
if ((hook == NULL)
|| NG_HOOK_NOT_VALID(hook)
|| NG_NODE_NOT_VALID(node) ) {
error = EIO;
NG_FREE_ITEM(item);
break;
}
/*
* If no receive method, just silently drop it.
* Give preference to the hook over-ride method
*/
if ((!(rcvdata = hook->hk_rcvdata))
&& (!(rcvdata = NG_HOOK_NODE(hook)->nd_type->rcvdata))) {
error = 0;
NG_FREE_ITEM(item);
break;
}
error = (*rcvdata)(hook, item);
break;
case NGQF_MESG:
if (hook) {
if (NG_HOOK_NOT_VALID(hook)) {
/*
* The hook has been zapped then we can't
* use it. Immediatly drop its reference.
* The message may not need it.
*/
NG_HOOK_UNREF(hook);
hook = NULL;
}
}
/*
* Similarly, if the node is a zombie there is
* nothing we can do with it, drop everything.
*/
if (NG_NODE_NOT_VALID(node)) {
TRAP_ERROR();
error = EINVAL;
NG_FREE_ITEM(item);
} else {
/*
* Call the appropriate message handler for the object.
* It is up to the message handler to free the message.
* If it's a generic message, handle it generically,
* otherwise call the type's message handler
* (if it exists)
* XXX (race). Remember that a queued message may
* reference a node or hook that has just been
* invalidated. It will exist as the queue code
* is holding a reference, but..
*/
struct ng_mesg *msg = NGI_MSG(item);
/*
* check if the generic handler owns it.
*/
if ((msg->header.typecookie == NGM_GENERIC_COOKIE)
&& ((msg->header.flags & NGF_RESP) == 0)) {
error = ng_generic_msg(node, item, hook);
break;
}
/*
* Now see if there is a handler (hook or node specific)
* in the target node. If none, silently discard.
*/
if (((!hook) || (!(rcvmsg = hook->hk_rcvmsg)))
&& (!(rcvmsg = node->nd_type->rcvmsg))) {
TRAP_ERROR();
error = 0;
NG_FREE_ITEM(item);
break;
}
error = (*rcvmsg)(node, item, hook);
}
break;
case NGQF_FN:
/*
* We have to implicitly trust the hook,
* as some of these are used for system purposes
* where the hook is invalid. In the case of
* the shutdown message we allow it to hit
* even if the node is invalid.
*/
if ((NG_NODE_NOT_VALID(node))
&& (NGI_FN(item) != &ng_rmnode)) {
TRAP_ERROR();
error = EINVAL;
NG_FREE_ITEM(item);
break;
}
(*NGI_FN(item))(node, hook, NGI_ARG1(item), NGI_ARG2(item));
NG_FREE_ITEM(item);
break;
}
/*
* We held references on some of the resources
* that we took from the item. Now that we have
* finished doing everything, drop those references.
*/
if (hook) {
NG_HOOK_UNREF(hook);
}
if (was_reader) {
ng_leave_read(&node->nd_input_queue);
} else {
ng_leave_write(&node->nd_input_queue);
}
/* Apply callback. */
if (apply != NULL)
(*apply)(context, error);
return (error);
}
/***********************************************************************
* Implement the 'generic' control messages
***********************************************************************/
static int
ng_generic_msg(node_p here, item_p item, hook_p lasthook)
{
int error = 0;
struct ng_mesg *msg;
struct ng_mesg *resp = NULL;
NGI_GET_MSG(item, msg);
if (msg->header.typecookie != NGM_GENERIC_COOKIE) {
TRAP_ERROR();
error = EINVAL;
goto out;
}
switch (msg->header.cmd) {
case NGM_SHUTDOWN:
ng_rmnode(here, NULL, NULL, 0);
break;
case NGM_MKPEER:
{
struct ngm_mkpeer *const mkp = (struct ngm_mkpeer *) msg->data;
if (msg->header.arglen != sizeof(*mkp)) {
TRAP_ERROR();
error = EINVAL;
break;
}
mkp->type[sizeof(mkp->type) - 1] = '\0';
mkp->ourhook[sizeof(mkp->ourhook) - 1] = '\0';
mkp->peerhook[sizeof(mkp->peerhook) - 1] = '\0';
error = ng_mkpeer(here, mkp->ourhook, mkp->peerhook, mkp->type);
break;
}
case NGM_CONNECT:
{
struct ngm_connect *const con =
(struct ngm_connect *) msg->data;
node_p node2;
if (msg->header.arglen != sizeof(*con)) {
TRAP_ERROR();
error = EINVAL;
break;
}
con->path[sizeof(con->path) - 1] = '\0';
con->ourhook[sizeof(con->ourhook) - 1] = '\0';
con->peerhook[sizeof(con->peerhook) - 1] = '\0';
/* Don't forget we get a reference.. */
error = ng_path2noderef(here, con->path, &node2, NULL);
if (error)
break;
error = ng_con_nodes(here, con->ourhook, node2, con->peerhook);
NG_NODE_UNREF(node2);
break;
}
case NGM_NAME:
{
struct ngm_name *const nam = (struct ngm_name *) msg->data;
if (msg->header.arglen != sizeof(*nam)) {
TRAP_ERROR();
error = EINVAL;
break;
}
nam->name[sizeof(nam->name) - 1] = '\0';
error = ng_name_node(here, nam->name);
break;
}
case NGM_RMHOOK:
{
struct ngm_rmhook *const rmh = (struct ngm_rmhook *) msg->data;
hook_p hook;
if (msg->header.arglen != sizeof(*rmh)) {
TRAP_ERROR();
error = EINVAL;
break;
}
rmh->ourhook[sizeof(rmh->ourhook) - 1] = '\0';
if ((hook = ng_findhook(here, rmh->ourhook)) != NULL)
ng_destroy_hook(hook);
break;
}
case NGM_NODEINFO:
{
struct nodeinfo *ni;
NG_MKRESPONSE(resp, msg, sizeof(*ni), M_NOWAIT);
if (resp == NULL) {
error = ENOMEM;
break;
}
/* Fill in node info */
ni = (struct nodeinfo *) resp->data;
if (NG_NODE_HAS_NAME(here))
strcpy(ni->name, NG_NODE_NAME(here));
strcpy(ni->type, here->nd_type->name);
ni->id = ng_node2ID(here);
ni->hooks = here->nd_numhooks;
break;
}
case NGM_LISTHOOKS:
{
const int nhooks = here->nd_numhooks;
struct hooklist *hl;
struct nodeinfo *ni;
hook_p hook;
/* Get response struct */
NG_MKRESPONSE(resp, msg, sizeof(*hl)
+ (nhooks * sizeof(struct linkinfo)), M_NOWAIT);
if (resp == NULL) {
error = ENOMEM;
break;
}
hl = (struct hooklist *) resp->data;
ni = &hl->nodeinfo;
/* Fill in node info */
if (NG_NODE_HAS_NAME(here))
strcpy(ni->name, NG_NODE_NAME(here));
strcpy(ni->type, here->nd_type->name);
ni->id = ng_node2ID(here);
/* Cycle through the linked list of hooks */
ni->hooks = 0;
LIST_FOREACH(hook, &here->nd_hooks, hk_hooks) {
struct linkinfo *const link = &hl->link[ni->hooks];
if (ni->hooks >= nhooks) {
log(LOG_ERR, "%s: number of %s changed\n",
__func__, "hooks");
break;
}
if (NG_HOOK_NOT_VALID(hook))
continue;
strcpy(link->ourhook, NG_HOOK_NAME(hook));
strcpy(link->peerhook, NG_PEER_HOOK_NAME(hook));
if (NG_PEER_NODE_NAME(hook)[0] != '\0')
strcpy(link->nodeinfo.name,
NG_PEER_NODE_NAME(hook));
strcpy(link->nodeinfo.type,
NG_PEER_NODE(hook)->nd_type->name);
link->nodeinfo.id = ng_node2ID(NG_PEER_NODE(hook));
link->nodeinfo.hooks = NG_PEER_NODE(hook)->nd_numhooks;
ni->hooks++;
}
break;
}
case NGM_LISTNAMES:
case NGM_LISTNODES:
{
const int unnamed = (msg->header.cmd == NGM_LISTNODES);
struct namelist *nl;
node_p node;
int num = 0;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx);
/* Count number of nodes */
LIST_FOREACH(node, &ng_nodelist, nd_nodes) {
if (NG_NODE_IS_VALID(node)
&& (unnamed || NG_NODE_HAS_NAME(node))) {
num++;
}
}
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
/* Get response struct */
NG_MKRESPONSE(resp, msg, sizeof(*nl)
+ (num * sizeof(struct nodeinfo)), M_NOWAIT);
if (resp == NULL) {
error = ENOMEM;
break;
}
nl = (struct namelist *) resp->data;
/* Cycle through the linked list of nodes */
nl->numnames = 0;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_nodelist_mtx);
LIST_FOREACH(node, &ng_nodelist, nd_nodes) {
struct nodeinfo *const np = &nl->nodeinfo[nl->numnames];
if (nl->numnames >= num) {
log(LOG_ERR, "%s: number of %s changed\n",
__func__, "nodes");
break;
}
if (NG_NODE_NOT_VALID(node))
continue;
if (!unnamed && (! NG_NODE_HAS_NAME(node)))
continue;
if (NG_NODE_HAS_NAME(node))
strcpy(np->name, NG_NODE_NAME(node));
strcpy(np->type, node->nd_type->name);
np->id = ng_node2ID(node);
np->hooks = node->nd_numhooks;
nl->numnames++;
}
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_nodelist_mtx);
break;
}
case NGM_LISTTYPES:
{
struct typelist *tl;
struct ng_type *type;
int num = 0;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_typelist_mtx);
/* Count number of types */
LIST_FOREACH(type, &ng_typelist, types) {
num++;
}
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_typelist_mtx);
/* Get response struct */
NG_MKRESPONSE(resp, msg, sizeof(*tl)
+ (num * sizeof(struct typeinfo)), M_NOWAIT);
if (resp == NULL) {
error = ENOMEM;
break;
}
tl = (struct typelist *) resp->data;
/* Cycle through the linked list of types */
tl->numtypes = 0;
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_typelist_mtx);
LIST_FOREACH(type, &ng_typelist, types) {
struct typeinfo *const tp = &tl->typeinfo[tl->numtypes];
if (tl->numtypes >= num) {
log(LOG_ERR, "%s: number of %s changed\n",
__func__, "types");
break;
}
strcpy(tp->type_name, type->name);
tp->numnodes = type->refs - 1; /* don't count list */
tl->numtypes++;
}
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_typelist_mtx);
break;
}
case NGM_BINARY2ASCII:
{
int bufSize = 20 * 1024; /* XXX hard coded constant */
const struct ng_parse_type *argstype;
const struct ng_cmdlist *c;
struct ng_mesg *binary, *ascii;
/* Data area must contain a valid netgraph message */
binary = (struct ng_mesg *)msg->data;
if (msg->header.arglen < sizeof(struct ng_mesg)
|| (msg->header.arglen - sizeof(struct ng_mesg)
< binary->header.arglen)) {
TRAP_ERROR();
error = EINVAL;
break;
}
/* Get a response message with lots of room */
NG_MKRESPONSE(resp, msg, sizeof(*ascii) + bufSize, M_NOWAIT);
if (resp == NULL) {
error = ENOMEM;
break;
}
ascii = (struct ng_mesg *)resp->data;
/* Copy binary message header to response message payload */
bcopy(binary, ascii, sizeof(*binary));
/* Find command by matching typecookie and command number */
for (c = here->nd_type->cmdlist;
c != NULL && c->name != NULL; c++) {
if (binary->header.typecookie == c->cookie
&& binary->header.cmd == c->cmd)
break;
}
if (c == NULL || c->name == NULL) {
for (c = ng_generic_cmds; c->name != NULL; c++) {
if (binary->header.typecookie == c->cookie
&& binary->header.cmd == c->cmd)
break;
}
if (c->name == NULL) {
NG_FREE_MSG(resp);
error = ENOSYS;
break;
}
}
/* Convert command name to ASCII */
snprintf(ascii->header.cmdstr, sizeof(ascii->header.cmdstr),
"%s", c->name);
/* Convert command arguments to ASCII */
argstype = (binary->header.flags & NGF_RESP) ?
c->respType : c->mesgType;
if (argstype == NULL) {
*ascii->data = '\0';
} else {
if ((error = ng_unparse(argstype,
(u_char *)binary->data,
ascii->data, bufSize)) != 0) {
NG_FREE_MSG(resp);
break;
}
}
/* Return the result as struct ng_mesg plus ASCII string */
bufSize = strlen(ascii->data) + 1;
ascii->header.arglen = bufSize;
resp->header.arglen = sizeof(*ascii) + bufSize;
break;
}
case NGM_ASCII2BINARY:
{
int bufSize = 2000; /* XXX hard coded constant */
const struct ng_cmdlist *c;
const struct ng_parse_type *argstype;
struct ng_mesg *ascii, *binary;
int off = 0;
/* Data area must contain at least a struct ng_mesg + '\0' */
ascii = (struct ng_mesg *)msg->data;
if ((msg->header.arglen < sizeof(*ascii) + 1)
|| (ascii->header.arglen < 1)
|| (msg->header.arglen
< sizeof(*ascii) + ascii->header.arglen)) {
TRAP_ERROR();
error = EINVAL;
break;
}
ascii->data[ascii->header.arglen - 1] = '\0';
/* Get a response message with lots of room */
NG_MKRESPONSE(resp, msg, sizeof(*binary) + bufSize, M_NOWAIT);
if (resp == NULL) {
error = ENOMEM;
break;
}
binary = (struct ng_mesg *)resp->data;
/* Copy ASCII message header to response message payload */
bcopy(ascii, binary, sizeof(*ascii));
/* Find command by matching ASCII command string */
for (c = here->nd_type->cmdlist;
c != NULL && c->name != NULL; c++) {
if (strcmp(ascii->header.cmdstr, c->name) == 0)
break;
}
if (c == NULL || c->name == NULL) {
for (c = ng_generic_cmds; c->name != NULL; c++) {
if (strcmp(ascii->header.cmdstr, c->name) == 0)
break;
}
if (c->name == NULL) {
NG_FREE_MSG(resp);
error = ENOSYS;
break;
}
}
/* Convert command name to binary */
binary->header.cmd = c->cmd;
binary->header.typecookie = c->cookie;
/* Convert command arguments to binary */
argstype = (binary->header.flags & NGF_RESP) ?
c->respType : c->mesgType;
if (argstype == NULL) {
bufSize = 0;
} else {
if ((error = ng_parse(argstype, ascii->data,
&off, (u_char *)binary->data, &bufSize)) != 0) {
NG_FREE_MSG(resp);
break;
}
}
/* Return the result */
binary->header.arglen = bufSize;
resp->header.arglen = sizeof(*binary) + bufSize;
break;
}
case NGM_TEXT_CONFIG:
case NGM_TEXT_STATUS:
/*
* This one is tricky as it passes the command down to the
* actual node, even though it is a generic type command.
* This means we must assume that the item/msg is already freed
* when control passes back to us.
*/
if (here->nd_type->rcvmsg != NULL) {
NGI_MSG(item) = msg; /* put it back as we found it */
return((*here->nd_type->rcvmsg)(here, item, lasthook));
}
/* Fall through if rcvmsg not supported */
default:
TRAP_ERROR();
error = EINVAL;
}
/*
* Sometimes a generic message may be statically allocated
* to avoid problems with allocating when in tight memeory situations.
* Don't free it if it is so.
* I break them appart here, because erros may cause a free if the item
* in which case we'd be doing it twice.
* they are kept together above, to simplify freeing.
*/
out:
NG_RESPOND_MSG(error, here, item, resp);
if (msg)
NG_FREE_MSG(msg);
return (error);
}
/************************************************************************
Queue element get/free routines
************************************************************************/
uma_zone_t ng_qzone;
static int maxalloc = 512; /* limit the damage of a leak */
TUNABLE_INT("net.graph.maxalloc", &maxalloc);
SYSCTL_INT(_net_graph, OID_AUTO, maxalloc, CTLFLAG_RDTUN, &maxalloc,
0, "Maximum number of queue items to allocate");
#ifdef NETGRAPH_DEBUG
static TAILQ_HEAD(, ng_item) ng_itemlist = TAILQ_HEAD_INITIALIZER(ng_itemlist);
static int allocated; /* number of items malloc'd */
#endif
/*
* Get a queue entry.
* This is usually called when a packet first enters netgraph.
* By definition, this is usually from an interrupt, or from a user.
* Users are not so important, but try be quick for the times that it's
* an interrupt.
*/
static __inline item_p
ng_getqblk(int flags)
{
item_p item = NULL;
int wait;
wait = (flags & NG_WAITOK) ? M_WAITOK : M_NOWAIT;
item = uma_zalloc(ng_qzone, wait | M_ZERO);
#ifdef NETGRAPH_DEBUG
if (item) {
mtx_lock(&ngq_mtx);
TAILQ_INSERT_TAIL(&ng_itemlist, item, all);
allocated++;
mtx_unlock(&ngq_mtx);
}
#endif
return (item);
}
/*
* Release a queue entry
*/
void
ng_free_item(item_p item)
{
/*
* The item may hold resources on it's own. We need to free
* these before we can free the item. What they are depends upon
* what kind of item it is. it is important that nodes zero
* out pointers to resources that they remove from the item
* or we release them again here.
*/
switch (item->el_flags & NGQF_TYPE) {
case NGQF_DATA:
/* If we have an mbuf still attached.. */
NG_FREE_M(_NGI_M(item));
break;
case NGQF_MESG:
_NGI_RETADDR(item) = 0;
NG_FREE_MSG(_NGI_MSG(item));
break;
case NGQF_FN:
/* nothing to free really, */
_NGI_FN(item) = NULL;
_NGI_ARG1(item) = NULL;
_NGI_ARG2(item) = 0;
case NGQF_UNDEF:
break;
}
/* If we still have a node or hook referenced... */
_NGI_CLR_NODE(item);
_NGI_CLR_HOOK(item);
#ifdef NETGRAPH_DEBUG
mtx_lock(&ngq_mtx);
TAILQ_REMOVE(&ng_itemlist, item, all);
allocated--;
mtx_unlock(&ngq_mtx);
#endif
uma_zfree(ng_qzone, item);
}
/************************************************************************
Module routines
************************************************************************/
/*
* Handle the loading/unloading of a netgraph node type module
*/
int
ng_mod_event(module_t mod, int event, void *data)
{
struct ng_type *const type = data;
int s, error = 0;
switch (event) {
case MOD_LOAD:
/* Register new netgraph node type */
s = splnet();
if ((error = ng_newtype(type)) != 0) {
splx(s);
break;
}
/* Call type specific code */
if (type->mod_event != NULL)
if ((error = (*type->mod_event)(mod, event, data))) {
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_typelist_mtx);
type->refs--; /* undo it */
LIST_REMOVE(type, types);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_typelist_mtx);
}
splx(s);
break;
case MOD_UNLOAD:
s = splnet();
if (type->refs > 1) { /* make sure no nodes exist! */
error = EBUSY;
} else {
if (type->refs == 0) {
/* failed load, nothing to undo */
splx(s);
break;
}
if (type->mod_event != NULL) { /* check with type */
error = (*type->mod_event)(mod, event, data);
if (error != 0) { /* type refuses.. */
splx(s);
break;
}
}
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock(&ng_typelist_mtx);
LIST_REMOVE(type, types);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock(&ng_typelist_mtx);
}
splx(s);
break;
default:
if (type->mod_event != NULL)
error = (*type->mod_event)(mod, event, data);
else
error = EOPNOTSUPP; /* XXX ? */
break;
}
return (error);
}
/*
* Handle loading and unloading for this code.
* The only thing we need to link into is the NETISR strucure.
*/
static int
ngb_mod_event(module_t mod, int event, void *data)
{
int error = 0;
switch (event) {
case MOD_LOAD:
/* Initialize everything. */
mtx_init(&ng_worklist_mtx, "ng_worklist", NULL, MTX_SPIN);
mtx_init(&ng_typelist_mtx, "netgraph types mutex", NULL,
MTX_DEF);
mtx_init(&ng_nodelist_mtx, "netgraph nodelist mutex", NULL,
MTX_DEF);
mtx_init(&ng_idhash_mtx, "netgraph idhash mutex", NULL,
MTX_DEF);
#ifdef NETGRAPH_DEBUG
mtx_init(&ngq_mtx, "netgraph item list mutex", NULL,
MTX_DEF);
#endif
ng_qzone = uma_zcreate("NetGraph items", sizeof(struct ng_item),
NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
uma_zone_set_max(ng_qzone, maxalloc);
netisr_register(NETISR_NETGRAPH, (netisr_t *)ngintr, NULL,
NETISR_MPSAFE);
break;
case MOD_UNLOAD:
/* You cant unload it because an interface may be using it. */
error = EBUSY;
break;
default:
error = EOPNOTSUPP;
break;
}
return (error);
}
static moduledata_t netgraph_mod = {
"netgraph",
ngb_mod_event,
(NULL)
};
DECLARE_MODULE(netgraph, netgraph_mod, SI_SUB_NETGRAPH, SI_ORDER_MIDDLE);
SYSCTL_NODE(_net, OID_AUTO, graph, CTLFLAG_RW, 0, "netgraph Family");
SYSCTL_INT(_net_graph, OID_AUTO, abi_version, CTLFLAG_RD, 0, NG_ABI_VERSION,"");
SYSCTL_INT(_net_graph, OID_AUTO, msg_version, CTLFLAG_RD, 0, NG_VERSION, "");
#ifdef NETGRAPH_DEBUG
void
dumphook (hook_p hook, char *file, int line)
{
printf("hook: name %s, %d refs, Last touched:\n",
_NG_HOOK_NAME(hook), hook->hk_refs);
printf(" Last active @ %s, line %d\n",
hook->lastfile, hook->lastline);
if (line) {
printf(" problem discovered at file %s, line %d\n", file, line);
}
}
void
dumpnode(node_p node, char *file, int line)
{
printf("node: ID [%x]: type '%s', %d hooks, flags 0x%x, %d refs, %s:\n",
_NG_NODE_ID(node), node->nd_type->name,
node->nd_numhooks, node->nd_flags,
node->nd_refs, node->nd_name);
printf(" Last active @ %s, line %d\n",
node->lastfile, node->lastline);
if (line) {
printf(" problem discovered at file %s, line %d\n", file, line);
}
}
void
dumpitem(item_p item, char *file, int line)
{
printf(" ACTIVE item, last used at %s, line %d",
item->lastfile, item->lastline);
switch(item->el_flags & NGQF_TYPE) {
case NGQF_DATA:
printf(" - [data]\n");
break;
case NGQF_MESG:
printf(" - retaddr[%d]:\n", _NGI_RETADDR(item));
break;
case NGQF_FN:
printf(" - fn@%p (%p, %p, %p, %d (%x))\n",
item->body.fn.fn_fn,
NGI_NODE(item),
NGI_HOOK(item),
item->body.fn.fn_arg1,
item->body.fn.fn_arg2,
item->body.fn.fn_arg2);
break;
case NGQF_UNDEF:
printf(" - UNDEFINED!\n");
}
if (line) {
printf(" problem discovered at file %s, line %d\n", file, line);
if (NGI_NODE(item)) {
printf("node %p ([%x])\n",
NGI_NODE(item), ng_node2ID(NGI_NODE(item)));
}
}
}
static void
ng_dumpitems(void)
{
item_p item;
int i = 1;
TAILQ_FOREACH(item, &ng_itemlist, all) {
printf("[%d] ", i++);
dumpitem(item, NULL, 0);
}
}
static void
ng_dumpnodes(void)
{
node_p node;
int i = 1;
mtx_lock(&ng_nodelist_mtx);
SLIST_FOREACH(node, &ng_allnodes, nd_all) {
printf("[%d] ", i++);
dumpnode(node, NULL, 0);
}
mtx_unlock(&ng_nodelist_mtx);
}
static void
ng_dumphooks(void)
{
hook_p hook;
int i = 1;
mtx_lock(&ng_nodelist_mtx);
SLIST_FOREACH(hook, &ng_allhooks, hk_all) {
printf("[%d] ", i++);
dumphook(hook, NULL, 0);
}
mtx_unlock(&ng_nodelist_mtx);
}
static int
sysctl_debug_ng_dump_items(SYSCTL_HANDLER_ARGS)
{
int error;
int val;
int i;
val = allocated;
i = 1;
error = sysctl_handle_int(oidp, &val, sizeof(int), req);
if (error != 0 || req->newptr == NULL)
return (error);
if (val == 42) {
ng_dumpitems();
ng_dumpnodes();
ng_dumphooks();
}
return (0);
}
SYSCTL_PROC(_debug, OID_AUTO, ng_dump_items, CTLTYPE_INT | CTLFLAG_RW,
0, sizeof(int), sysctl_debug_ng_dump_items, "I", "Number of allocated items");
#endif /* NETGRAPH_DEBUG */
/***********************************************************************
* Worklist routines
**********************************************************************/
/* NETISR thread enters here */
/*
* Pick a node off the list of nodes with work,
* try get an item to process off it.
* If there are no more, remove the node from the list.
*/
static void
ngintr(void)
{
item_p item;
node_p node = NULL;
for (;;) {
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock_spin(&ng_worklist_mtx);
node = TAILQ_FIRST(&ng_worklist);
if (!node) {
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin(&ng_worklist_mtx);
break;
}
node->nd_flags &= ~NGF_WORKQ;
TAILQ_REMOVE(&ng_worklist, node, nd_work);
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin(&ng_worklist_mtx);
/*
* We have the node. We also take over the reference
* that the list had on it.
* Now process as much as you can, until it won't
* let you have another item off the queue.
* All this time, keep the reference
* that lets us be sure that the node still exists.
* Let the reference go at the last minute.
* ng_dequeue will put us back on the worklist
* if there is more too do. This may be of use if there
* are Multiple Processors and multiple Net threads in the
* future.
*/
for (;;) {
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock_spin(&node->nd_input_queue.q_mtx);
item = ng_dequeue(&node->nd_input_queue);
if (item == NULL) {
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin(&node->nd_input_queue.q_mtx);
break; /* go look for another node */
} else {
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_unlock_spin(&node->nd_input_queue.q_mtx);
NGI_GET_NODE(item, node); /* zaps stored node */
ng_apply_item(node, item);
NG_NODE_UNREF(node);
}
}
NG_NODE_UNREF(node);
}
}
static void
ng_worklist_remove(node_p node)
{
Change and clean the mutex lock interface. mtx_enter(lock, type) becomes: mtx_lock(lock) for sleep locks (MTX_DEF-initialized locks) mtx_lock_spin(lock) for spin locks (MTX_SPIN-initialized) similarily, for releasing a lock, we now have: mtx_unlock(lock) for MTX_DEF and mtx_unlock_spin(lock) for MTX_SPIN. We change the caller interface for the two different types of locks because the semantics are entirely different for each case, and this makes it explicitly clear and, at the same time, it rids us of the extra `type' argument. The enter->lock and exit->unlock change has been made with the idea that we're "locking data" and not "entering locked code" in mind. Further, remove all additional "flags" previously passed to the lock acquire/release routines with the exception of two: MTX_QUIET and MTX_NOSWITCH The functionality of these flags is preserved and they can be passed to the lock/unlock routines by calling the corresponding wrappers: mtx_{lock, unlock}_flags(lock, flag(s)) and mtx_{lock, unlock}_spin_flags(lock, flag(s)) for MTX_DEF and MTX_SPIN locks, respectively. Re-inline some lock acq/rel code; in the sleep lock case, we only inline the _obtain_lock()s in order to ensure that the inlined code fits into a cache line. In the spin lock case, we inline recursion and actually only perform a function call if we need to spin. This change has been made with the idea that we generally tend to avoid spin locks and that also the spin locks that we do have and are heavily used (i.e. sched_lock) do recurse, and therefore in an effort to reduce function call overhead for some architectures (such as alpha), we inline recursion for this case. Create a new malloc type for the witness code and retire from using the M_DEV type. The new type is called M_WITNESS and is only declared if WITNESS is enabled. Begin cleaning up some machdep/mutex.h code - specifically updated the "optimized" inlined code in alpha/mutex.h and wrote MTX_LOCK_SPIN and MTX_UNLOCK_SPIN asm macros for the i386/mutex.h as we presently need those. Finally, caught up to the interface changes in all sys code. Contributors: jake, jhb, jasone (in no particular order)
2001-02-09 06:11:45 +00:00
mtx_lock_spin(&ng_worklist_mtx);
if (node->nd_flags & NGF_WORKQ) {
node->nd_flags &= ~NGF_WORKQ;
TAILQ_REMOVE(&ng_worklist, node, nd_work);
mtx_unlock_spin(&ng_worklist_mtx);
NG_NODE_UNREF(node);
} else {
mtx_unlock_spin(&ng_worklist_mtx);
}
}
/*
* XXX
* It's posible that a debugging NG_NODE_REF may need
* to be outside the mutex zone
*/
static void
ng_setisr(node_p node)
{
Problem description: At the end of ng_snd_item(), node queue is processed. In certain netgraph setups deep recursive calls can occur. For example this happens, when two nodes are connected and can send items to each other in both directions. If, for some reason, both nodes have a lot of items in their queues, then the processing thread will recurse between these two nodes, delivering items left and right, going deeper in the stack. Other setups can suffer from deep recursion, too. The following factors can influence risk of deep netgraph call: - periodical write-access events on node - combination of slow link and fast one in one graph - net.inet.ip.fastforwarding Changes made: - In ng_acquire_{read,write}() do not dequeue another item. Instead, call ng_setisr() for this node. - At the end of ng_snd_item(), do not process queue. Call ng_setisr(), if there are any dequeueable items on node queue. - In ng_setisr() narrow worklist mutex holding. - In ng_setisr() assert queue mutex. Theoretically, the first two changes should negatively affect performance. To check this, some profiling was made: 1) In general real tasks, no noticable performance difference was found. 2) The following test was made: two multithreaded nodes and one single-threaded were connected into a ring. A large queues of packets were sent around this ring. Time to pass the ring N times was measured. This is a very vacuous test: no items/mbufs are allocated, no upcalls or downcalls outside of netgraph. It doesn't represent a real load, it is a stress test for ng_acquire_{read,write}() and item queueing functions. Surprisingly, the performance impact was positive! New code is 13% faster on UP and 17% faster on SMP, in this particular test. The problem was originally found, described, analyzed and original patch was written by Roselyn Lee from Vernier Networks. Thanks! Submitted by: Roselyn Lee <rosel verniernetworks com>
2005-07-21 12:08:37 +00:00
mtx_assert(&node->nd_input_queue.q_mtx, MA_OWNED);
Problem description: At the end of ng_snd_item(), node queue is processed. In certain netgraph setups deep recursive calls can occur. For example this happens, when two nodes are connected and can send items to each other in both directions. If, for some reason, both nodes have a lot of items in their queues, then the processing thread will recurse between these two nodes, delivering items left and right, going deeper in the stack. Other setups can suffer from deep recursion, too. The following factors can influence risk of deep netgraph call: - periodical write-access events on node - combination of slow link and fast one in one graph - net.inet.ip.fastforwarding Changes made: - In ng_acquire_{read,write}() do not dequeue another item. Instead, call ng_setisr() for this node. - At the end of ng_snd_item(), do not process queue. Call ng_setisr(), if there are any dequeueable items on node queue. - In ng_setisr() narrow worklist mutex holding. - In ng_setisr() assert queue mutex. Theoretically, the first two changes should negatively affect performance. To check this, some profiling was made: 1) In general real tasks, no noticable performance difference was found. 2) The following test was made: two multithreaded nodes and one single-threaded were connected into a ring. A large queues of packets were sent around this ring. Time to pass the ring N times was measured. This is a very vacuous test: no items/mbufs are allocated, no upcalls or downcalls outside of netgraph. It doesn't represent a real load, it is a stress test for ng_acquire_{read,write}() and item queueing functions. Surprisingly, the performance impact was positive! New code is 13% faster on UP and 17% faster on SMP, in this particular test. The problem was originally found, described, analyzed and original patch was written by Roselyn Lee from Vernier Networks. Thanks! Submitted by: Roselyn Lee <rosel verniernetworks com>
2005-07-21 12:08:37 +00:00
if ((node->nd_flags & NGF_WORKQ) == 0) {
/*
* If we are not already on the work queue,
* then put us on.
*/
node->nd_flags |= NGF_WORKQ;
Problem description: At the end of ng_snd_item(), node queue is processed. In certain netgraph setups deep recursive calls can occur. For example this happens, when two nodes are connected and can send items to each other in both directions. If, for some reason, both nodes have a lot of items in their queues, then the processing thread will recurse between these two nodes, delivering items left and right, going deeper in the stack. Other setups can suffer from deep recursion, too. The following factors can influence risk of deep netgraph call: - periodical write-access events on node - combination of slow link and fast one in one graph - net.inet.ip.fastforwarding Changes made: - In ng_acquire_{read,write}() do not dequeue another item. Instead, call ng_setisr() for this node. - At the end of ng_snd_item(), do not process queue. Call ng_setisr(), if there are any dequeueable items on node queue. - In ng_setisr() narrow worklist mutex holding. - In ng_setisr() assert queue mutex. Theoretically, the first two changes should negatively affect performance. To check this, some profiling was made: 1) In general real tasks, no noticable performance difference was found. 2) The following test was made: two multithreaded nodes and one single-threaded were connected into a ring. A large queues of packets were sent around this ring. Time to pass the ring N times was measured. This is a very vacuous test: no items/mbufs are allocated, no upcalls or downcalls outside of netgraph. It doesn't represent a real load, it is a stress test for ng_acquire_{read,write}() and item queueing functions. Surprisingly, the performance impact was positive! New code is 13% faster on UP and 17% faster on SMP, in this particular test. The problem was originally found, described, analyzed and original patch was written by Roselyn Lee from Vernier Networks. Thanks! Submitted by: Roselyn Lee <rosel verniernetworks com>
2005-07-21 12:08:37 +00:00
mtx_lock_spin(&ng_worklist_mtx);
TAILQ_INSERT_TAIL(&ng_worklist, node, nd_work);
Problem description: At the end of ng_snd_item(), node queue is processed. In certain netgraph setups deep recursive calls can occur. For example this happens, when two nodes are connected and can send items to each other in both directions. If, for some reason, both nodes have a lot of items in their queues, then the processing thread will recurse between these two nodes, delivering items left and right, going deeper in the stack. Other setups can suffer from deep recursion, too. The following factors can influence risk of deep netgraph call: - periodical write-access events on node - combination of slow link and fast one in one graph - net.inet.ip.fastforwarding Changes made: - In ng_acquire_{read,write}() do not dequeue another item. Instead, call ng_setisr() for this node. - At the end of ng_snd_item(), do not process queue. Call ng_setisr(), if there are any dequeueable items on node queue. - In ng_setisr() narrow worklist mutex holding. - In ng_setisr() assert queue mutex. Theoretically, the first two changes should negatively affect performance. To check this, some profiling was made: 1) In general real tasks, no noticable performance difference was found. 2) The following test was made: two multithreaded nodes and one single-threaded were connected into a ring. A large queues of packets were sent around this ring. Time to pass the ring N times was measured. This is a very vacuous test: no items/mbufs are allocated, no upcalls or downcalls outside of netgraph. It doesn't represent a real load, it is a stress test for ng_acquire_{read,write}() and item queueing functions. Surprisingly, the performance impact was positive! New code is 13% faster on UP and 17% faster on SMP, in this particular test. The problem was originally found, described, analyzed and original patch was written by Roselyn Lee from Vernier Networks. Thanks! Submitted by: Roselyn Lee <rosel verniernetworks com>
2005-07-21 12:08:37 +00:00
mtx_unlock_spin(&ng_worklist_mtx);
NG_NODE_REF(node); /* XXX fafe in mutex? */
}
schednetisr(NETISR_NETGRAPH);
}
/***********************************************************************
* Externally useable functions to set up a queue item ready for sending
***********************************************************************/
#ifdef NETGRAPH_DEBUG
#define ITEM_DEBUG_CHECKS \
do { \
if (NGI_NODE(item) ) { \
printf("item already has node"); \
kdb_enter("has node"); \
NGI_CLR_NODE(item); \
} \
if (NGI_HOOK(item) ) { \
printf("item already has hook"); \
kdb_enter("has hook"); \
NGI_CLR_HOOK(item); \
} \
} while (0)
#else
#define ITEM_DEBUG_CHECKS
#endif
/*
* Put mbuf into the item.
* Hook and node references will be removed when the item is dequeued.
* (or equivalent)
* (XXX) Unsafe because no reference held by peer on remote node.
* remote node might go away in this timescale.
* We know the hooks can't go away because that would require getting
* a writer item on both nodes and we must have at least a reader
* here to eb able to do this.
* Note that the hook loaded is the REMOTE hook.
*
* This is possibly in the critical path for new data.
*/
item_p
ng_package_data(struct mbuf *m, int flags)
{
item_p item;
if ((item = ng_getqblk(flags)) == NULL) {
NG_FREE_M(m);
return (NULL);
}
ITEM_DEBUG_CHECKS;
item->el_flags = NGQF_DATA;
item->el_next = NULL;
NGI_M(item) = m;
return (item);
}
/*
* Allocate a queue item and put items into it..
* Evaluate the address as this will be needed to queue it and
* to work out what some of the fields should be.
* Hook and node references will be removed when the item is dequeued.
* (or equivalent)
*/
item_p
ng_package_msg(struct ng_mesg *msg, int flags)
{
item_p item;
if ((item = ng_getqblk(flags)) == NULL) {
NG_FREE_MSG(msg);
return (NULL);
}
ITEM_DEBUG_CHECKS;
item->el_flags = NGQF_MESG;
item->el_next = NULL;
/*
* Set the current lasthook into the queue item
*/
NGI_MSG(item) = msg;
NGI_RETADDR(item) = 0;
return (item);
}
#define SET_RETADDR(item, here, retaddr) \
do { /* Data or fn items don't have retaddrs */ \
if ((item->el_flags & NGQF_TYPE) == NGQF_MESG) { \
if (retaddr) { \
NGI_RETADDR(item) = retaddr; \
} else { \
/* \
* The old return address should be ok. \
* If there isn't one, use the address \
* here. \
*/ \
if (NGI_RETADDR(item) == 0) { \
NGI_RETADDR(item) \
= ng_node2ID(here); \
} \
} \
} \
} while (0)
int
ng_address_hook(node_p here, item_p item, hook_p hook, ng_ID_t retaddr)
{
hook_p peer;
node_p peernode;
ITEM_DEBUG_CHECKS;
/*
* Quick sanity check..
* Since a hook holds a reference on it's node, once we know
* that the peer is still connected (even if invalid,) we know
* that the peer node is present, though maybe invalid.
*/
if ((hook == NULL)
|| NG_HOOK_NOT_VALID(hook)
|| (NG_HOOK_PEER(hook) == NULL)
|| NG_HOOK_NOT_VALID(NG_HOOK_PEER(hook))
|| NG_NODE_NOT_VALID(NG_PEER_NODE(hook))) {
NG_FREE_ITEM(item);
TRAP_ERROR();
return (ENETDOWN);
}
/*
* Transfer our interest to the other (peer) end.
*/
peer = NG_HOOK_PEER(hook);
NG_HOOK_REF(peer);
NGI_SET_HOOK(item, peer);
peernode = NG_PEER_NODE(hook);
NG_NODE_REF(peernode);
NGI_SET_NODE(item, peernode);
SET_RETADDR(item, here, retaddr);
return (0);
}
int
ng_address_path(node_p here, item_p item, char *address, ng_ID_t retaddr)
{
node_p dest = NULL;
hook_p hook = NULL;
int error;
ITEM_DEBUG_CHECKS;
/*
* Note that ng_path2noderef increments the reference count
* on the node for us if it finds one. So we don't have to.
*/
error = ng_path2noderef(here, address, &dest, &hook);
if (error) {
NG_FREE_ITEM(item);
return (error);
}
NGI_SET_NODE(item, dest);
if ( hook) {
NG_HOOK_REF(hook); /* don't let it go while on the queue */
NGI_SET_HOOK(item, hook);
}
SET_RETADDR(item, here, retaddr);
return (0);
}
int
ng_address_ID(node_p here, item_p item, ng_ID_t ID, ng_ID_t retaddr)
{
node_p dest;
ITEM_DEBUG_CHECKS;
/*
* Find the target node.
*/
dest = ng_ID2noderef(ID); /* GETS REFERENCE! */
if (dest == NULL) {
NG_FREE_ITEM(item);
TRAP_ERROR();
return(EINVAL);
}
/* Fill out the contents */
item->el_flags = NGQF_MESG;
item->el_next = NULL;
NGI_SET_NODE(item, dest);
NGI_CLR_HOOK(item);
SET_RETADDR(item, here, retaddr);
return (0);
}
/*
* special case to send a message to self (e.g. destroy node)
* Possibly indicate an arrival hook too.
* Useful for removing that hook :-)
*/
item_p
ng_package_msg_self(node_p here, hook_p hook, struct ng_mesg *msg)
{
item_p item;
/*
* Find the target node.
* If there is a HOOK argument, then use that in preference
* to the address.
*/
if ((item = ng_getqblk(NG_NOFLAGS)) == NULL) {
NG_FREE_MSG(msg);
return (NULL);
}
/* Fill out the contents */
item->el_flags = NGQF_MESG;
item->el_next = NULL;
NG_NODE_REF(here);
NGI_SET_NODE(item, here);
if (hook) {
NG_HOOK_REF(hook);
NGI_SET_HOOK(item, hook);
}
NGI_MSG(item) = msg;
NGI_RETADDR(item) = ng_node2ID(here);
return (item);
}
int
ng_send_fn1(node_p node, hook_p hook, ng_item_fn *fn, void * arg1, int arg2,
int flags)
{
item_p item;
if ((item = ng_getqblk(flags)) == NULL) {
return (ENOMEM);
}
item->el_flags = NGQF_FN | NGQF_WRITER;
NG_NODE_REF(node); /* and one for the item */
NGI_SET_NODE(item, node);
if (hook) {
NG_HOOK_REF(hook);
NGI_SET_HOOK(item, hook);
}
NGI_FN(item) = fn;
NGI_ARG1(item) = arg1;
NGI_ARG2(item) = arg2;
return(ng_snd_item(item, flags));
}
/*
* Official timeout routines for Netgraph nodes.
*/
static void
ng_callout_trampoline(void *arg)
{
item_p item = arg;
ng_snd_item(item, 0);
}
int
ng_callout(struct callout *c, node_p node, hook_p hook, int ticks,
ng_item_fn *fn, void * arg1, int arg2)
{
item_p item;
if ((item = ng_getqblk(NG_NOFLAGS)) == NULL)
return (ENOMEM);
item->el_flags = NGQF_FN | NGQF_WRITER;
NG_NODE_REF(node); /* and one for the item */
NGI_SET_NODE(item, node);
if (hook) {
NG_HOOK_REF(hook);
NGI_SET_HOOK(item, hook);
}
NGI_FN(item) = fn;
NGI_ARG1(item) = arg1;
NGI_ARG2(item) = arg2;
callout_reset(c, ticks, &ng_callout_trampoline, item);
return (0);
}
/* A special modified version of untimeout() */
int
ng_uncallout(struct callout *c, node_p node)
{
item_p item;
int rval;
KASSERT(c != NULL, ("ng_uncallout: NULL callout"));
KASSERT(node != NULL, ("ng_uncallout: NULL node"));
rval = callout_stop(c);
item = c->c_arg;
/* Do an extra check */
if ((rval > 0) && (c->c_func == &ng_callout_trampoline) &&
(NGI_NODE(item) == node)) {
/*
* We successfully removed it from the queue before it ran
* So now we need to unreference everything that was
* given extra references. (NG_FREE_ITEM does this).
*/
NG_FREE_ITEM(item);
}
return (rval);
}
/*
* Set the address, if none given, give the node here.
*/
void
ng_replace_retaddr(node_p here, item_p item, ng_ID_t retaddr)
{
if (retaddr) {
NGI_RETADDR(item) = retaddr;
} else {
/*
* The old return address should be ok.
* If there isn't one, use the address here.
*/
NGI_RETADDR(item) = ng_node2ID(here);
}
}
#define TESTING
#ifdef TESTING
/* just test all the macros */
void
ng_macro_test(item_p item);
void
ng_macro_test(item_p item)
{
node_p node = NULL;
hook_p hook = NULL;
struct mbuf *m;
struct ng_mesg *msg;
ng_ID_t retaddr;
int error;
NGI_GET_M(item, m);
NGI_GET_MSG(item, msg);
retaddr = NGI_RETADDR(item);
NG_SEND_DATA(error, hook, m, NULL);
NG_SEND_DATA_ONLY(error, hook, m);
NG_FWD_NEW_DATA(error, item, hook, m);
NG_FWD_ITEM_HOOK(error, item, hook);
NG_SEND_MSG_HOOK(error, node, msg, hook, retaddr);
NG_SEND_MSG_ID(error, node, msg, retaddr, retaddr);
NG_SEND_MSG_PATH(error, node, msg, ".:", retaddr);
NG_FWD_MSG_HOOK(error, node, item, hook, retaddr);
}
#endif /* TESTING */