2 * Copyright (c) 2003, 2004 Matthew Dillon. All rights reserved.
3 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
4 * Copyright (c) 2003 Jonathan Lemon. All rights reserved.
5 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved.
7 * This code is derived from software contributed to The DragonFly Project
8 * by Jonathan Lemon, Jeffrey M. Hsu, and Matthew Dillon.
10 * Jonathan Lemon gave Jeffrey Hsu permission to combine his copyright
11 * into this one around July 8 2004.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. Neither the name of The DragonFly Project nor the names of its
22 * contributors may be used to endorse or promote products derived
23 * from this software without specific, prior written permission.
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
28 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
29 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
30 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
31 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
32 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
33 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
34 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
35 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * $DragonFly: src/sys/net/netisr.c,v 1.49 2008/11/01 10:29:31 sephe Exp $
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/msgport.h>
47 #include <sys/interrupt.h>
48 #include <sys/socket.h>
49 #include <sys/sysctl.h>
50 #include <sys/socketvar.h>
52 #include <net/if_var.h>
53 #include <net/netisr.h>
54 #include <machine/cpufunc.h>
56 #include <sys/thread2.h>
57 #include <sys/msgport2.h>
58 #include <net/netmsg2.h>
59 #include <sys/mplock2.h>
61 #define NETISR_GET_MPLOCK(ni) \
63 if (((ni)->ni_flags & NETISR_FLAG_MPSAFE) == 0) \
67 #define NETISR_REL_MPLOCK(ni) \
69 if (((ni)->ni_flags & NETISR_FLAG_MPSAFE) == 0) \
73 static void netmsg_sync_func(struct netmsg
*msg
);
75 struct netmsg_port_registration
{
76 TAILQ_ENTRY(netmsg_port_registration
) npr_entry
;
80 static struct netisr netisrs
[NETISR_MAX
];
81 static TAILQ_HEAD(,netmsg_port_registration
) netreglist
;
83 /* Per-CPU thread to handle any protocol. */
84 struct thread netisr_cpu
[MAXCPU
];
85 lwkt_port netisr_afree_rport
;
86 lwkt_port netisr_adone_rport
;
87 lwkt_port netisr_apanic_rport
;
88 lwkt_port netisr_sync_port
;
90 static int (*netmsg_fwd_port_fn
)(lwkt_port_t
, lwkt_msg_t
);
92 static int netisr_mpsafe_thread
= NETMSG_SERVICE_ADAPTIVE
;
93 TUNABLE_INT("net.netisr.mpsafe_thread", &netisr_mpsafe_thread
);
95 SYSCTL_NODE(_net
, OID_AUTO
, netisr
, CTLFLAG_RW
, 0, "netisr");
96 SYSCTL_INT(_net_netisr
, OID_AUTO
, mpsafe_thread
, CTLFLAG_RW
,
97 &netisr_mpsafe_thread
, 0,
98 "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)");
101 NETISR_TO_MSGF(const struct netisr
*ni
)
105 if (ni
->ni_flags
& NETISR_FLAG_MPSAFE
)
106 msg_flags
|= MSGF_MPSAFE
;
111 * netisr_afree_rport replymsg function, only used to handle async
112 * messages which the sender has abandoned to their fate.
115 netisr_autofree_reply(lwkt_port_t port
, lwkt_msg_t msg
)
117 kfree(msg
, M_LWKTMSG
);
121 * We need a custom putport function to handle the case where the
122 * message target is the current thread's message port. This case
123 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
124 * then turns around and executes a network operation synchronously.
126 * To prevent deadlocking, we must execute these self-referential messages
127 * synchronously, effectively turning the message into a glorified direct
128 * procedure call back into the protocol stack. The operation must be
129 * complete on return or we will deadlock, so panic if it isn't.
132 netmsg_put_port(lwkt_port_t port
, lwkt_msg_t lmsg
)
134 netmsg_t netmsg
= (void *)lmsg
;
136 if ((lmsg
->ms_flags
& MSGF_SYNC
) && port
== &curthread
->td_msgport
) {
137 netmsg
->nm_dispatch(netmsg
);
138 if ((lmsg
->ms_flags
& MSGF_DONE
) == 0)
139 panic("netmsg_put_port: self-referential deadlock on netport");
142 return(netmsg_fwd_port_fn(port
, lmsg
));
147 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
148 * because they depend on the user proc context for a number of things
149 * (like creds) which we have not yet incorporated into the message structure.
151 * However, we maintain or message/port abstraction. Having a special
152 * synchronous port which runs the commands synchronously gives us the
153 * ability to serialize operations in one place later on when we start
157 netmsg_sync_putport(lwkt_port_t port
, lwkt_msg_t lmsg
)
159 netmsg_t netmsg
= (void *)lmsg
;
161 KKASSERT((lmsg
->ms_flags
& MSGF_DONE
) == 0);
163 lmsg
->ms_target_port
= port
; /* required for abort */
164 netmsg
->nm_dispatch(netmsg
);
173 TAILQ_INIT(&netreglist
);
176 * Create default per-cpu threads for generic protocol handling.
178 for (i
= 0; i
< ncpus
; ++i
) {
179 lwkt_create(netmsg_service_loop
, &netisr_mpsafe_thread
, NULL
,
180 &netisr_cpu
[i
], TDF_NETWORK
| TDF_MPSAFE
, i
,
182 netmsg_service_port_init(&netisr_cpu
[i
].td_msgport
);
186 * The netisr_afree_rport is a special reply port which automatically
187 * frees the replied message. The netisr_adone_rport simply marks
188 * the message as being done. The netisr_apanic_rport panics if
189 * the message is replied to.
191 lwkt_initport_replyonly(&netisr_afree_rport
, netisr_autofree_reply
);
192 lwkt_initport_replyonly_null(&netisr_adone_rport
);
193 lwkt_initport_panic(&netisr_apanic_rport
);
196 * The netisr_syncport is a special port which executes the message
197 * synchronously and waits for it if EASYNC is returned.
199 lwkt_initport_putonly(&netisr_sync_port
, netmsg_sync_putport
);
202 SYSINIT(netisr
, SI_SUB_PRE_DRIVERS
, SI_ORDER_FIRST
, netisr_init
, NULL
);
205 * Finish initializing the message port for a netmsg service. This also
206 * registers the port for synchronous cleanup operations such as when an
207 * ifnet is being destroyed. There is no deregistration API yet.
210 netmsg_service_port_init(lwkt_port_t port
)
212 struct netmsg_port_registration
*reg
;
215 * Override the putport function. Our custom function checks for
216 * self-references and executes such commands synchronously.
218 if (netmsg_fwd_port_fn
== NULL
)
219 netmsg_fwd_port_fn
= port
->mp_putport
;
220 KKASSERT(netmsg_fwd_port_fn
== port
->mp_putport
);
221 port
->mp_putport
= netmsg_put_port
;
224 * Keep track of ports using the netmsg API so we can synchronize
225 * certain operations (such as freeing an ifnet structure) across all
228 reg
= kmalloc(sizeof(*reg
), M_TEMP
, M_WAITOK
|M_ZERO
);
229 reg
->npr_port
= port
;
230 TAILQ_INSERT_TAIL(&netreglist
, reg
, npr_entry
);
234 * This function synchronizes the caller with all netmsg services. For
235 * example, if an interface is being removed we must make sure that all
236 * packets related to that interface complete processing before the structure
237 * can actually be freed. This sort of synchronization is an alternative to
238 * ref-counting the netif, removing the ref counting overhead in favor of
239 * placing additional overhead in the netif freeing sequence (where it is
243 netmsg_service_sync(void)
245 struct netmsg_port_registration
*reg
;
248 netmsg_init(&smsg
, NULL
, &curthread
->td_msgport
,
249 MSGF_MPSAFE
, netmsg_sync_func
);
251 TAILQ_FOREACH(reg
, &netreglist
, npr_entry
) {
252 lwkt_domsg(reg
->npr_port
, &smsg
.nm_lmsg
, 0);
257 * The netmsg function simply replies the message. API semantics require
258 * EASYNC to be returned if the netmsg function disposes of the message.
261 netmsg_sync_func(struct netmsg
*msg
)
263 lwkt_replymsg(&msg
->nm_lmsg
, 0);
267 * Service a netmsg request and modify the BGL lock state if appropriate.
268 * The new BGL lock state is returned (1:locked, 0:unlocked).
271 netmsg_service(struct netmsg
*msg
, int mpsafe_mode
, int mplocked
)
274 * If nm_so is non-NULL the message is related to a socket. Sockets
275 * can migrate between protocol processing threads when they connect,
276 * due to an implied connect during a sendmsg(), or when a connection
279 * If this occurs any messages already queued to the original thread
280 * or which race the change must be forwarded to the new protocol
283 * MPSAFE - socket changes are synchronous to the current protocol port
284 * so if the port can only change out from under us if it is
285 * already different from the current port anyway so we forward
286 * it. It is possible to chase a changing port, which is fine.
288 if (msg
->nm_so
&& msg
->nm_so
->so_port
!= &curthread
->td_msgport
) {
289 lwkt_forwardmsg(msg
->nm_so
->so_port
, &msg
->nm_lmsg
);
294 * Adjust the mplock dynamically.
296 switch (mpsafe_mode
) {
297 case NETMSG_SERVICE_ADAPTIVE
: /* Adaptive BGL */
298 if (msg
->nm_lmsg
.ms_flags
& MSGF_MPSAFE
) {
303 msg
->nm_dispatch(msg
);
304 /* Leave mpunlocked */
308 /* mplocked = 1; not needed */
310 msg
->nm_dispatch(msg
);
313 /* Leave mpunlocked, next msg might be mpsafe */
317 case NETMSG_SERVICE_MPSAFE
: /* No BGL */
322 msg
->nm_dispatch(msg
);
323 /* Leave mpunlocked */
331 msg
->nm_dispatch(msg
);
339 * Generic netmsg service loop. Some protocols may roll their own but all
340 * must do the basic command dispatch function call done here.
343 netmsg_service_loop(void *arg
)
346 int mplocked
, *mpsafe_mode
= arg
;
349 * Thread was started with TDF_MPSAFE
356 while ((msg
= lwkt_waitport(&curthread
->td_msgport
, 0))) {
357 mplocked
= netmsg_service(msg
, *mpsafe_mode
, mplocked
);
362 * Call the netisr directly.
363 * Queueing may be done in the msg port layer at its discretion.
366 netisr_dispatch(int num
, struct mbuf
*m
)
368 /* just queue it for now XXX JH */
369 netisr_queue(num
, m
);
373 * Same as netisr_dispatch(), but always queue.
374 * This is either used in places where we are not confident that
375 * direct dispatch is possible, or where queueing is required.
378 netisr_queue(int num
, struct mbuf
*m
)
381 struct netmsg_packet
*pmsg
;
384 KASSERT((num
> 0 && num
<= (sizeof(netisrs
)/sizeof(netisrs
[0]))),
385 ("%s: bad isr %d", __func__
, num
));
388 if (ni
->ni_handler
== NULL
) {
389 kprintf("%s: unregistered isr %d\n", __func__
, num
);
394 if ((port
= ni
->ni_mport(&m
)) == NULL
)
397 pmsg
= &m
->m_hdr
.mh_netmsg
;
399 netmsg_init(&pmsg
->nm_netmsg
, NULL
, &netisr_apanic_rport
,
400 NETISR_TO_MSGF(ni
), ni
->ni_handler
);
402 pmsg
->nm_netmsg
.nm_lmsg
.u
.ms_result
= num
;
403 lwkt_sendmsg(port
, &pmsg
->nm_netmsg
.nm_lmsg
);
408 netisr_register(int num
, pkt_portfn_t mportfn
,
409 pktinfo_portfn_t mportfn_pktinfo
, netisr_fn_t handler
,
414 KASSERT((num
> 0 && num
<= (sizeof(netisrs
)/sizeof(netisrs
[0]))),
415 ("netisr_register: bad isr %d", num
));
418 ni
->ni_mport
= mportfn
;
419 ni
->ni_mport_pktinfo
= mportfn_pktinfo
;
420 ni
->ni_handler
= handler
;
421 ni
->ni_flags
= flags
;
422 netmsg_init(&ni
->ni_netmsg
, NULL
, &netisr_adone_rport
,
423 NETISR_TO_MSGF(ni
), NULL
);
427 netisr_unregister(int num
)
429 KASSERT((num
> 0 && num
<= (sizeof(netisrs
)/sizeof(netisrs
[0]))),
430 ("unregister_netisr: bad isr number: %d\n", num
));
437 * Return message port for default handler thread on CPU 0.
440 cpu0_portfn(struct mbuf
**mptr
)
442 struct mbuf
*m
= *mptr
;
445 m
->m_pkthdr
.hash
= cpu
;
446 m
->m_flags
|= M_HASH
;
447 return (&netisr_cpu
[cpu
].td_msgport
);
453 return (&netisr_cpu
[cpu
].td_msgport
);
457 * If the current thread is a network protocol thread (TDF_NETWORK),
458 * then return the current thread's message port.
459 * XXX Else, return the current CPU's netisr message port.
464 if (curthread
->td_flags
& TDF_NETWORK
)
465 return &curthread
->td_msgport
;
467 return cpu_portfn(mycpuid
);
472 cpu0_soport(struct socket
*so __unused
, struct sockaddr
*nam __unused
,
473 struct mbuf
**dummy __unused
)
475 return (&netisr_cpu
[0].td_msgport
);
479 cpu0_ctlport(int cmd __unused
, struct sockaddr
*sa __unused
,
480 void *extra __unused
)
482 return (&netisr_cpu
[0].td_msgport
);
486 sync_soport(struct socket
*so __unused
, struct sockaddr
*nam __unused
,
487 struct mbuf
**dummy __unused
)
489 return (&netisr_sync_port
);
493 * schednetisr() is used to call the netisr handler from the appropriate
494 * netisr thread for polling and other purposes.
496 * This function may be called from a hard interrupt or IPI and must be
497 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
498 * trying to allocate one. We must get ourselves onto the target cpu
499 * to safely check the MSGF_DONE bit on the message but since the message
500 * will be sent to that cpu anyway this does not add any extra work beyond
501 * what lwkt_sendmsg() would have already had to do to schedule the target
505 schednetisr_remote(void *data
)
507 int num
= (int)(intptr_t)data
;
508 struct netisr
*ni
= &netisrs
[num
];
509 lwkt_port_t port
= &netisr_cpu
[0].td_msgport
;
512 pmsg
= &netisrs
[num
].ni_netmsg
;
514 if (pmsg
->nm_lmsg
.ms_flags
& MSGF_DONE
) {
515 netmsg_init(pmsg
, NULL
, &netisr_adone_rport
,
516 NETISR_TO_MSGF(ni
), ni
->ni_handler
);
517 pmsg
->nm_lmsg
.u
.ms_result
= num
;
518 lwkt_sendmsg(port
, &pmsg
->nm_lmsg
);
526 KASSERT((num
> 0 && num
<= (sizeof(netisrs
)/sizeof(netisrs
[0]))),
527 ("schednetisr: bad isr %d", num
));
529 if (mycpu
->gd_cpuid
!= 0) {
530 lwkt_send_ipiq(globaldata_find(0),
531 schednetisr_remote
, (void *)(intptr_t)num
);
533 schednetisr_remote((void *)(intptr_t)num
);
536 schednetisr_remote((void *)(intptr_t)num
);
541 netisr_find_port(int num
, struct mbuf
**m0
)
545 struct mbuf
*m
= *m0
;
549 KASSERT((num
> 0 && num
<= (sizeof(netisrs
)/sizeof(netisrs
[0]))),
550 ("%s: bad isr %d", __func__
, num
));
553 if (ni
->ni_mport
== NULL
) {
554 kprintf("%s: unregistered isr %d\n", __func__
, num
);
559 if ((port
= ni
->ni_mport(&m
)) == NULL
)
567 netisr_run(int num
, struct mbuf
*m
)
570 struct netmsg_packet
*pmsg
;
572 KASSERT((num
> 0 && num
<= (sizeof(netisrs
)/sizeof(netisrs
[0]))),
573 ("%s: bad isr %d", __func__
, num
));
576 if (ni
->ni_handler
== NULL
) {
577 kprintf("%s: unregistered isr %d\n", __func__
, num
);
582 pmsg
= &m
->m_hdr
.mh_netmsg
;
584 netmsg_init(&pmsg
->nm_netmsg
, NULL
, &netisr_apanic_rport
,
587 pmsg
->nm_netmsg
.nm_lmsg
.u
.ms_result
= num
;
589 NETISR_GET_MPLOCK(ni
);
590 ni
->ni_handler(&pmsg
->nm_netmsg
);
591 NETISR_REL_MPLOCK(ni
);
595 pktinfo_portfn_cpu0(const struct pktinfo
*dummy __unused
,
598 m
->m_pkthdr
.hash
= 0;
599 return &netisr_cpu
[0].td_msgport
;
603 pktinfo_portfn_notsupp(const struct pktinfo
*dummy __unused
,
604 struct mbuf
*m __unused
)
610 netisr_find_pktinfo_port(const struct pktinfo
*pi
, struct mbuf
*m
)
613 int num
= pi
->pi_netisr
;
615 KASSERT(m
->m_flags
& M_HASH
, ("packet does not contain hash\n"));
616 KASSERT((num
> 0 && num
<= (sizeof(netisrs
)/sizeof(netisrs
[0]))),
617 ("%s: bad isr %d", __func__
, num
));
620 if (ni
->ni_mport_pktinfo
== NULL
) {
621 kprintf("%s: unregistered isr %d\n", __func__
, num
);
624 return ni
->ni_mport_pktinfo(pi
, m
);