freebsd-dev/sys/net/pfil.c

683 lines
17 KiB
C
Raw Normal View History

/* $FreeBSD$ */
/* $NetBSD: pfil.c,v 1.20 2001/11/12 23:49:46 lukem Exp $ */
/*-
* SPDX-License-Identifier: BSD-3-Clause
*
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
* Copyright (c) 2019 Gleb Smirnoff <glebius@FreeBSD.org>
* Copyright (c) 1996 Matthew R. Green
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. The name of the author may not be used to endorse or promote products
* derived from this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
* OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
* AND ON 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 ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
#include <sys/param.h>
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
#include <sys/conf.h>
#include <sys/kernel.h>
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
#include <sys/epoch.h>
#include <sys/errno.h>
Somewhat re-factor the read/write locking mechanism associated with the packet filtering mechanisms to use the new rwlock(9) locking API: - Drop the variables stored in the phil_head structure which were specific to conditions and the home rolled read/write locking mechanism. - Drop some includes which were used for condition variables - Drop the inline functions, and convert them to macros. Also, move these macros into pfil.h - Move pfil list locking macros intp phil.h as well - Rename ph_busy_count to ph_nhooks. This variable will represent the number of IN/OUT hooks registered with the pfil head structure - Define PFIL_HOOKED macro which evaluates to true if there are any hooks to be ran by pfil_run_hooks - In the IP/IP6 stacks, change the ph_busy_count comparison to use the new PFIL_HOOKED macro. - Drop optimization in pfil_run_hooks which checks to see if there are any hooks to be ran, and returns if not. This check is already performed by the IP stacks when they call: if (!PFIL_HOOKED(ph)) goto skip_hooks; - Drop in assertion which makes sure that the number of hooks never drops below 0 for good measure. This in theory should never happen, and if it does than there are problems somewhere - Drop special logic around PFIL_WAITOK because rw_wlock(9) does not sleep - Drop variables which support home rolled read/write locking mechanism from the IPFW firewall chain structure. - Swap out the read/write firewall chain lock internal to use the rwlock(9) API instead of our home rolled version - Convert the inlined functions to macros Reviewed by: mlaier, andre, glebius Thanks to: jhb for the new locking API
2006-02-02 03:13:16 +00:00
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/socket.h>
#include <sys/socketvar.h>
#include <sys/systm.h>
#include <sys/lock.h>
#include <sys/mutex.h>
#include <sys/proc.h>
#include <sys/queue.h>
#include <sys/ucred.h>
#include <sys/jail.h>
#include <net/if.h>
#include <net/if_var.h>
#include <net/pfil.h>
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
static MALLOC_DEFINE(M_PFIL, "pfil", "pfil(9) packet filter hooks");
static int pfil_ioctl(struct cdev *, u_long, caddr_t, int, struct thread *);
static struct cdevsw pfil_cdevsw = {
.d_ioctl = pfil_ioctl,
.d_name = PFILDEV,
.d_version = D_VERSION,
};
static struct cdev *pfil_dev;
static struct mtx pfil_lock;
MTX_SYSINIT(pfil_mtxinit, &pfil_lock, "pfil(9) lock", MTX_DEF);
#define PFIL_LOCK() mtx_lock(&pfil_lock)
#define PFIL_UNLOCK() mtx_unlock(&pfil_lock)
#define PFIL_LOCK_ASSERT() mtx_assert(&pfil_lock, MA_OWNED)
#define PFIL_EPOCH net_epoch_preempt
#define PFIL_EPOCH_ENTER(et) epoch_enter_preempt(net_epoch_preempt, &(et))
#define PFIL_EPOCH_EXIT(et) epoch_exit_preempt(net_epoch_preempt, &(et))
struct pfil_hook {
pfil_func_t hook_func;
void *hook_ruleset;
int hook_flags;
int hook_links;
enum pfil_types hook_type;
const char *hook_modname;
const char *hook_rulname;
LIST_ENTRY(pfil_hook) hook_list;
};
struct pfil_link {
CK_STAILQ_ENTRY(pfil_link) link_chain;
pfil_func_t link_func;
void *link_ruleset;
int link_flags;
struct pfil_hook *link_hook;
struct epoch_context link_epoch_ctx;
};
typedef CK_STAILQ_HEAD(pfil_chain, pfil_link) pfil_chain_t;
struct pfil_head {
int head_nhooksin;
int head_nhooksout;
pfil_chain_t head_in;
pfil_chain_t head_out;
int head_flags;
enum pfil_types head_type;
LIST_ENTRY(pfil_head) head_list;
const char *head_name;
};
LIST_HEAD(pfilheadhead, pfil_head);
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
VNET_DEFINE_STATIC(struct pfilheadhead, pfil_head_list) =
LIST_HEAD_INITIALIZER(pfil_head_list);
#define V_pfil_head_list VNET(pfil_head_list)
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
LIST_HEAD(pfilhookhead, pfil_hook);
VNET_DEFINE_STATIC(struct pfilhookhead, pfil_hook_list) =
LIST_HEAD_INITIALIZER(pfil_hook_list);
#define V_pfil_hook_list VNET(pfil_hook_list)
static struct pfil_link *pfil_link_remove(pfil_chain_t *, pfil_hook_t );
static void pfil_link_free(epoch_context_t);
int
pfil_realloc(pfil_packet_t *p, int flags, struct ifnet *ifp)
{
struct mbuf *m;
MPASS(flags & PFIL_MEMPTR);
if ((m = m_devget(p->mem, PFIL_LENGTH(flags), 0, ifp, NULL)) == NULL)
return (ENOMEM);
*p = pfil_packet_align(*p);
*p->m = m;
return (0);
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
static __noinline int
pfil_fake_mbuf(pfil_func_t func, pfil_packet_t *p, struct ifnet *ifp, int flags,
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
void *ruleset, struct inpcb *inp)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct mbuf m, *mp;
pfil_return_t rv;
(void)m_init(&m, M_NOWAIT, MT_DATA, M_NOFREE | M_PKTHDR);
m_extadd(&m, p->mem, PFIL_LENGTH(flags), NULL, NULL, NULL, 0,
EXT_RXRING);
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
m.m_len = m.m_pkthdr.len = PFIL_LENGTH(flags);
mp = &m;
flags &= ~(PFIL_MEMPTR | PFIL_LENMASK);
rv = func(&mp, ifp, flags, ruleset, inp);
if (rv == PFIL_PASS && mp != &m) {
/*
* Firewalls that need pfil_fake_mbuf() most likely don't
* know they need return PFIL_REALLOCED.
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
*/
rv = PFIL_REALLOCED;
*p = pfil_packet_align(*p);
*p->m = mp;
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
return (rv);
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
/*
* pfil_run_hooks() runs the specified packet filter hook chain.
*/
int
pfil_run_hooks(struct pfil_head *head, pfil_packet_t p, struct ifnet *ifp,
int flags, struct inpcb *inp)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct epoch_tracker et;
pfil_chain_t *pch;
struct pfil_link *link;
pfil_return_t rv;
bool realloc = false;
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
if (PFIL_DIR(flags) == PFIL_IN)
pch = &head->head_in;
else if (__predict_true(PFIL_DIR(flags) == PFIL_OUT))
pch = &head->head_out;
else
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
panic("%s: bogus flags %d", __func__, flags);
rv = PFIL_PASS;
PFIL_EPOCH_ENTER(et);
CK_STAILQ_FOREACH(link, pch, link_chain) {
if ((flags & PFIL_MEMPTR) && !(link->link_flags & PFIL_MEMPTR))
rv = pfil_fake_mbuf(link->link_func, &p, ifp, flags,
link->link_ruleset, inp);
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
else
rv = (*link->link_func)(p, ifp, flags,
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
link->link_ruleset, inp);
if (rv == PFIL_DROPPED || rv == PFIL_CONSUMED)
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
break;
else if (rv == PFIL_REALLOCED) {
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
flags &= ~(PFIL_MEMPTR | PFIL_LENMASK);
realloc = true;
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
}
}
PFIL_EPOCH_EXIT(et);
if (realloc && rv == PFIL_PASS)
rv = PFIL_REALLOCED;
return (rv);
}
/*
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
* pfil_head_register() registers a pfil_head with the packet filter hook
* mechanism.
*/
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
pfil_head_t
pfil_head_register(struct pfil_head_args *pa)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_head *head, *list;
MPASS(pa->pa_version == PFIL_VERSION);
head = malloc(sizeof(struct pfil_head), M_PFIL, M_WAITOK);
2013-03-19 05:51:47 +00:00
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
head->head_nhooksin = head->head_nhooksout = 0;
head->head_flags = pa->pa_flags;
head->head_type = pa->pa_type;
head->head_name = pa->pa_headname;
CK_STAILQ_INIT(&head->head_in);
CK_STAILQ_INIT(&head->head_out);
PFIL_LOCK();
LIST_FOREACH(list, &V_pfil_head_list, head_list)
if (strcmp(pa->pa_headname, list->head_name) == 0) {
printf("pfil: duplicate head \"%s\"\n",
pa->pa_headname);
}
LIST_INSERT_HEAD(&V_pfil_head_list, head, head_list);
PFIL_UNLOCK();
return (head);
}
/*
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
* pfil_head_unregister() removes a pfil_head from the packet filter hook
* mechanism. The producer of the hook promises that all outstanding
* invocations of the hook have completed before it unregisters the hook.
*/
void
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
pfil_head_unregister(pfil_head_t ph)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_link *link, *next;
PFIL_LOCK();
LIST_REMOVE(ph, head_list);
2013-03-19 05:51:47 +00:00
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
CK_STAILQ_FOREACH_SAFE(link, &ph->head_in, link_chain, next) {
link->link_hook->hook_links--;
free(link, M_PFIL);
}
CK_STAILQ_FOREACH_SAFE(link, &ph->head_out, link_chain, next) {
link->link_hook->hook_links--;
free(link, M_PFIL);
}
PFIL_UNLOCK();
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
pfil_hook_t
pfil_add_hook(struct pfil_hook_args *pa)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_hook *hook, *list;
MPASS(pa->pa_version == PFIL_VERSION);
hook = malloc(sizeof(struct pfil_hook), M_PFIL, M_WAITOK | M_ZERO);
hook->hook_func = pa->pa_func;
hook->hook_ruleset = pa->pa_ruleset;
hook->hook_flags = pa->pa_flags;
hook->hook_type = pa->pa_type;
hook->hook_modname = pa->pa_modname;
hook->hook_rulname = pa->pa_rulname;
PFIL_LOCK();
LIST_FOREACH(list, &V_pfil_hook_list, hook_list)
if (strcmp(pa->pa_modname, list->hook_modname) == 0 &&
strcmp(pa->pa_rulname, list->hook_rulname) == 0) {
printf("pfil: duplicate hook \"%s:%s\"\n",
pa->pa_modname, pa->pa_rulname);
}
LIST_INSERT_HEAD(&V_pfil_hook_list, hook, hook_list);
PFIL_UNLOCK();
2013-03-19 05:51:47 +00:00
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
return (hook);
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
static int
pfil_unlink(struct pfil_link_args *pa, pfil_head_t head, pfil_hook_t hook)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_link *in, *out;
PFIL_LOCK_ASSERT();
if (pa->pa_flags & PFIL_IN) {
in = pfil_link_remove(&head->head_in, hook);
if (in != NULL) {
head->head_nhooksin--;
hook->hook_links--;
}
} else
in = NULL;
if (pa->pa_flags & PFIL_OUT) {
out = pfil_link_remove(&head->head_out, hook);
if (out != NULL) {
head->head_nhooksout--;
hook->hook_links--;
}
} else
out = NULL;
PFIL_UNLOCK();
if (in != NULL)
epoch_call(PFIL_EPOCH, &in->link_epoch_ctx, pfil_link_free);
if (out != NULL)
epoch_call(PFIL_EPOCH, &out->link_epoch_ctx, pfil_link_free);
2013-03-19 05:51:47 +00:00
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
if (in == NULL && out == NULL)
return (ENOENT);
else
return (0);
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
int
pfil_link(struct pfil_link_args *pa)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_link *in, *out, *link;
struct pfil_head *head;
struct pfil_hook *hook;
int error;
MPASS(pa->pa_version == PFIL_VERSION);
if ((pa->pa_flags & (PFIL_IN | PFIL_UNLINK)) == PFIL_IN)
in = malloc(sizeof(*in), M_PFIL, M_WAITOK | M_ZERO);
else
in = NULL;
if ((pa->pa_flags & (PFIL_OUT | PFIL_UNLINK)) == PFIL_OUT)
out = malloc(sizeof(*out), M_PFIL, M_WAITOK | M_ZERO);
else
out = NULL;
PFIL_LOCK();
if (pa->pa_flags & PFIL_HEADPTR)
head = pa->pa_head;
else
LIST_FOREACH(head, &V_pfil_head_list, head_list)
if (strcmp(pa->pa_headname, head->head_name) == 0)
break;
if (pa->pa_flags & PFIL_HOOKPTR)
hook = pa->pa_hook;
else
LIST_FOREACH(hook, &V_pfil_hook_list, hook_list)
if (strcmp(pa->pa_modname, hook->hook_modname) == 0 &&
strcmp(pa->pa_rulname, hook->hook_rulname) == 0)
break;
if (head == NULL || hook == NULL) {
error = ENOENT;
goto fail;
}
if (pa->pa_flags & PFIL_UNLINK)
return (pfil_unlink(pa, head, hook));
if (head->head_type != hook->hook_type ||
((hook->hook_flags & pa->pa_flags) & ~head->head_flags)) {
error = EINVAL;
goto fail;
}
if (pa->pa_flags & PFIL_IN)
CK_STAILQ_FOREACH(link, &head->head_in, link_chain)
if (link->link_hook == hook) {
error = EEXIST;
goto fail;
}
if (pa->pa_flags & PFIL_OUT)
CK_STAILQ_FOREACH(link, &head->head_out, link_chain)
if (link->link_hook == hook) {
error = EEXIST;
goto fail;
}
if (pa->pa_flags & PFIL_IN) {
in->link_hook = hook;
in->link_func = hook->hook_func;
in->link_flags = hook->hook_flags;
in->link_ruleset = hook->hook_ruleset;
if (pa->pa_flags & PFIL_APPEND)
CK_STAILQ_INSERT_TAIL(&head->head_in, in, link_chain);
else
CK_STAILQ_INSERT_HEAD(&head->head_in, in, link_chain);
hook->hook_links++;
head->head_nhooksin++;
}
if (pa->pa_flags & PFIL_OUT) {
out->link_hook = hook;
out->link_func = hook->hook_func;
out->link_flags = hook->hook_flags;
out->link_ruleset = hook->hook_ruleset;
if (pa->pa_flags & PFIL_APPEND)
CK_STAILQ_INSERT_HEAD(&head->head_out, out, link_chain);
else
CK_STAILQ_INSERT_TAIL(&head->head_out, out, link_chain);
hook->hook_links++;
head->head_nhooksout++;
}
PFIL_UNLOCK();
return (0);
2013-03-19 05:51:47 +00:00
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
fail:
PFIL_UNLOCK();
free(in, M_PFIL);
free(out, M_PFIL);
return (error);
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
static void
pfil_link_free(epoch_context_t ctx)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_link *link;
2013-03-19 05:51:47 +00:00
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
link = __containerof(ctx, struct pfil_link, link_epoch_ctx);
free(link, M_PFIL);
}
/*
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
* pfil_remove_hook removes a filter from all filtering points.
*/
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
void
pfil_remove_hook(pfil_hook_t hook)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_head *head;
struct pfil_link *in, *out;
PFIL_LOCK();
LIST_FOREACH(head, &V_pfil_head_list, head_list) {
retry:
in = pfil_link_remove(&head->head_in, hook);
if (in != NULL) {
head->head_nhooksin--;
hook->hook_links--;
epoch_call(PFIL_EPOCH, &in->link_epoch_ctx,
pfil_link_free);
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
out = pfil_link_remove(&head->head_out, hook);
if (out != NULL) {
head->head_nhooksout--;
hook->hook_links--;
epoch_call(PFIL_EPOCH, &out->link_epoch_ctx,
pfil_link_free);
}
if (in != NULL || out != NULL)
/* What if some stupid admin put same filter twice? */
goto retry;
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
LIST_REMOVE(hook, hook_list);
PFIL_UNLOCK();
MPASS(hook->hook_links == 0);
free(hook, M_PFIL);
}
/*
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
* Internal: Remove a pfil hook from a hook chain.
*/
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
static struct pfil_link *
pfil_link_remove(pfil_chain_t *chain, pfil_hook_t hook)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_link *link;
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
PFIL_LOCK_ASSERT();
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
CK_STAILQ_FOREACH(link, chain, link_chain)
if (link->link_hook == hook) {
CK_STAILQ_REMOVE(chain, link, pfil_link, link_chain);
return (link);
}
return (NULL);
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
static void
pfil_init(const void *unused __unused)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct make_dev_args args;
int error;
make_dev_args_init(&args);
args.mda_flags = MAKEDEV_WAITOK | MAKEDEV_CHECKNAME;
args.mda_devsw = &pfil_cdevsw;
args.mda_uid = UID_ROOT;
args.mda_gid = GID_WHEEL;
args.mda_mode = 0600;
error = make_dev_s(&args, &pfil_dev, PFILDEV);
KASSERT(error == 0, ("%s: failed to create dev: %d", __func__, error));
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
/*
* Make sure the pfil bits are first before any possible subsystem which
* might piggyback on the SI_SUB_PROTO_PFIL.
*/
SYSINIT(pfil_init, SI_SUB_PROTO_PFIL, SI_ORDER_FIRST, pfil_init, NULL);
/*
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
* User control interface.
*/
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
static int pfilioc_listheads(struct pfilioc_list *);
static int pfilioc_listhooks(struct pfilioc_list *);
static int pfilioc_link(struct pfilioc_link *);
static int
pfil_ioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags,
struct thread *td)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
int error;
CURVNET_SET(TD_TO_VNET(td));
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
error = 0;
switch (cmd) {
case PFILIOC_LISTHEADS:
error = pfilioc_listheads((struct pfilioc_list *)addr);
break;
case PFILIOC_LISTHOOKS:
error = pfilioc_listhooks((struct pfilioc_list *)addr);
break;
case PFILIOC_LINK:
error = pfilioc_link((struct pfilioc_link *)addr);
break;
default:
error = EINVAL;
break;
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
}
CURVNET_RESTORE();
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
return (error);
}
static int
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
pfilioc_listheads(struct pfilioc_list *req)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_head *head;
struct pfil_link *link;
struct pfilioc_head *iohead;
struct pfilioc_hook *iohook;
u_int nheads, nhooks, hd, hk;
int error;
PFIL_LOCK();
restart:
nheads = nhooks = 0;
LIST_FOREACH(head, &V_pfil_head_list, head_list) {
nheads++;
nhooks += head->head_nhooksin + head->head_nhooksout;
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
PFIL_UNLOCK();
if (req->pio_nheads < nheads || req->pio_nhooks < nhooks) {
req->pio_nheads = nheads;
req->pio_nhooks = nhooks;
return (0);
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
iohead = malloc(sizeof(*iohead) * nheads, M_TEMP, M_WAITOK);
iohook = malloc(sizeof(*iohook) * nhooks, M_TEMP, M_WAITOK);
hd = hk = 0;
PFIL_LOCK();
LIST_FOREACH(head, &V_pfil_head_list, head_list) {
if (hd + 1 > nheads ||
hk + head->head_nhooksin + head->head_nhooksout > nhooks) {
/* Configuration changed during malloc(). */
free(iohead, M_TEMP);
free(iohook, M_TEMP);
goto restart;
}
strlcpy(iohead[hd].pio_name, head->head_name,
sizeof(iohead[0].pio_name));
iohead[hd].pio_nhooksin = head->head_nhooksin;
iohead[hd].pio_nhooksout = head->head_nhooksout;
iohead[hd].pio_type = head->head_type;
CK_STAILQ_FOREACH(link, &head->head_in, link_chain) {
strlcpy(iohook[hk].pio_module,
link->link_hook->hook_modname,
sizeof(iohook[0].pio_module));
strlcpy(iohook[hk].pio_ruleset,
link->link_hook->hook_rulname,
sizeof(iohook[0].pio_ruleset));
hk++;
}
CK_STAILQ_FOREACH(link, &head->head_out, link_chain) {
strlcpy(iohook[hk].pio_module,
link->link_hook->hook_modname,
sizeof(iohook[0].pio_module));
strlcpy(iohook[hk].pio_ruleset,
link->link_hook->hook_rulname,
sizeof(iohook[0].pio_ruleset));
hk++;
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
hd++;
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
PFIL_UNLOCK();
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
error = copyout(iohead, req->pio_heads,
sizeof(*iohead) * min(hd, req->pio_nheads));
if (error == 0)
error = copyout(iohook, req->pio_hooks,
sizeof(*iohook) * min(req->pio_nhooks, hk));
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
req->pio_nheads = hd;
req->pio_nhooks = hk;
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
free(iohead, M_TEMP);
free(iohook, M_TEMP);
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
return (error);
}
static int
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
pfilioc_listhooks(struct pfilioc_list *req)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_hook *hook;
struct pfilioc_hook *iohook;
u_int nhooks, hk;
int error;
PFIL_LOCK();
restart:
nhooks = 0;
LIST_FOREACH(hook, &V_pfil_hook_list, hook_list)
nhooks++;
PFIL_UNLOCK();
if (req->pio_nhooks < nhooks) {
req->pio_nhooks = nhooks;
return (0);
}
iohook = malloc(sizeof(*iohook) * nhooks, M_TEMP, M_WAITOK);
hk = 0;
PFIL_LOCK();
LIST_FOREACH(hook, &V_pfil_hook_list, hook_list) {
if (hk + 1 > nhooks) {
/* Configuration changed during malloc(). */
free(iohook, M_TEMP);
goto restart;
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
strlcpy(iohook[hk].pio_module, hook->hook_modname,
sizeof(iohook[0].pio_module));
strlcpy(iohook[hk].pio_ruleset, hook->hook_rulname,
sizeof(iohook[0].pio_ruleset));
iohook[hk].pio_type = hook->hook_type;
iohook[hk].pio_flags = hook->hook_flags;
hk++;
}
PFIL_UNLOCK();
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
error = copyout(iohook, req->pio_hooks,
sizeof(*iohook) * min(req->pio_nhooks, hk));
req->pio_nhooks = hk;
free(iohook, M_TEMP);
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
return (error);
}
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
static int
pfilioc_link(struct pfilioc_link *req)
{
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
struct pfil_link_args args;
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
if (req->pio_flags & ~(PFIL_IN | PFIL_OUT | PFIL_UNLINK | PFIL_APPEND))
return (EINVAL);
New pfil(9) KPI together with newborn pfil API and control utility. The KPI have been reviewed and cleansed of features that were planned back 20 years ago and never implemented. The pfil(9) internals have been made opaque to protocols with only returned types and function declarations exposed. The KPI is made more strict, but at the same time more extensible, as kernel uses same command structures that userland ioctl uses. In nutshell [KA]PI is about declaring filtering points, declaring filters and linking and unlinking them together. New [KA]PI makes it possible to reconfigure pfil(9) configuration: change order of hooks, rehook filter from one filtering point to a different one, disconnect a hook on output leaving it on input only, prepend/append a filter to existing list of filters. Now it possible for a single packet filter to provide multiple rulesets that may be linked to different points. Think of per-interface ACLs in Cisco or Juniper. None of existing packet filters yet support that, however limited usage is already possible, e.g. default ruleset can be moved to single interface, as soon as interface would pride their filtering points. Another future feature is possiblity to create pfil heads, that provide not an mbuf pointer but just a memory pointer with length. That would allow filtering at very early stages of a packet lifecycle, e.g. when packet has just been received by a NIC and no mbuf was yet allocated. Differential Revision: https://reviews.freebsd.org/D18951
2019-01-31 23:01:03 +00:00
args.pa_version = PFIL_VERSION;
args.pa_flags = req->pio_flags;
args.pa_headname = req->pio_name;
args.pa_modname = req->pio_module;
args.pa_rulname = req->pio_ruleset;
return (pfil_link(&args));
}