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
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/msgport.h>
45 #include <sys/interrupt.h>
46 #include <sys/socket.h>
47 #include <sys/sysctl.h>
48 #include <sys/socketvar.h>
50 #include <net/if_var.h>
51 #include <net/netisr2.h>
52 #include <machine/cpufunc.h>
53 #include <machine/smp.h>
55 #include <sys/thread2.h>
56 #include <sys/msgport2.h>
57 #include <net/netmsg2.h>
58 #include <sys/mplock2.h>
60 static void netmsg_service_loop(void *arg
);
61 static void netisr_hashfn0(struct mbuf
**mp
, int hoff
);
62 static void netisr_nohashck(struct mbuf
*, const struct pktinfo
*);
64 struct netmsg_port_registration
{
65 TAILQ_ENTRY(netmsg_port_registration
) npr_entry
;
69 struct netmsg_rollup
{
70 TAILQ_ENTRY(netmsg_rollup
) ru_entry
;
75 struct netmsg_barrier
{
76 struct netmsg_base base
;
77 volatile cpumask_t
*br_cpumask
;
78 volatile uint32_t br_done
;
81 #define NETISR_BR_NOTDONE 0x1
82 #define NETISR_BR_WAITDONE 0x80000000
84 struct netisr_barrier
{
85 struct netmsg_barrier
*br_msgs
[MAXCPU
];
89 static struct netisr netisrs
[NETISR_MAX
];
90 static TAILQ_HEAD(,netmsg_port_registration
) netreglist
;
91 static TAILQ_HEAD(,netmsg_rollup
) netrulist
;
93 /* Per-CPU thread to handle any protocol. */
94 struct thread netisr_cpu
[MAXCPU
];
95 lwkt_port netisr_afree_rport
;
96 lwkt_port netisr_afree_free_so_rport
;
97 lwkt_port netisr_adone_rport
;
98 lwkt_port netisr_apanic_rport
;
99 lwkt_port netisr_sync_port
;
101 static int (*netmsg_fwd_port_fn
)(lwkt_port_t
, lwkt_msg_t
);
103 SYSCTL_NODE(_net
, OID_AUTO
, netisr
, CTLFLAG_RW
, 0, "netisr");
104 static int netisr_rollup_limit
= 32;
105 SYSCTL_INT(_net_netisr
, OID_AUTO
, rollup_limit
, CTLFLAG_RW
,
106 &netisr_rollup_limit
, 0, "Message to process before rollup");
110 * netisr_afree_rport replymsg function, only used to handle async
111 * messages which the sender has abandoned to their fate.
114 netisr_autofree_reply(lwkt_port_t port
, lwkt_msg_t msg
)
116 kfree(msg
, M_LWKTMSG
);
120 netisr_autofree_free_so_reply(lwkt_port_t port
, lwkt_msg_t msg
)
122 sofree(((netmsg_t
)msg
)->base
.nm_so
);
123 kfree(msg
, M_LWKTMSG
);
127 * We need a custom putport function to handle the case where the
128 * message target is the current thread's message port. This case
129 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
130 * then turns around and executes a network operation synchronously.
132 * To prevent deadlocking, we must execute these self-referential messages
133 * synchronously, effectively turning the message into a glorified direct
134 * procedure call back into the protocol stack. The operation must be
135 * complete on return or we will deadlock, so panic if it isn't.
137 * However, the target function is under no obligation to immediately
138 * reply the message. It may forward it elsewhere.
141 netmsg_put_port(lwkt_port_t port
, lwkt_msg_t lmsg
)
143 netmsg_base_t nmsg
= (void *)lmsg
;
145 if ((lmsg
->ms_flags
& MSGF_SYNC
) && port
== &curthread
->td_msgport
) {
146 nmsg
->nm_dispatch((netmsg_t
)nmsg
);
149 return(netmsg_fwd_port_fn(port
, lmsg
));
154 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
155 * because they depend on the user proc context for a number of things
156 * (like creds) which we have not yet incorporated into the message structure.
158 * However, we maintain or message/port abstraction. Having a special
159 * synchronous port which runs the commands synchronously gives us the
160 * ability to serialize operations in one place later on when we start
164 netmsg_sync_putport(lwkt_port_t port
, lwkt_msg_t lmsg
)
166 netmsg_base_t nmsg
= (void *)lmsg
;
168 KKASSERT((lmsg
->ms_flags
& MSGF_DONE
) == 0);
170 lmsg
->ms_target_port
= port
; /* required for abort */
171 nmsg
->nm_dispatch((netmsg_t
)nmsg
);
180 TAILQ_INIT(&netreglist
);
181 TAILQ_INIT(&netrulist
);
184 * Create default per-cpu threads for generic protocol handling.
186 for (i
= 0; i
< ncpus
; ++i
) {
187 lwkt_create(netmsg_service_loop
, NULL
, NULL
,
189 TDF_NOSTART
|TDF_FORCE_SPINPORT
|TDF_FIXEDCPU
,
190 i
, "netisr_cpu %d", i
);
191 netmsg_service_port_init(&netisr_cpu
[i
].td_msgport
);
192 lwkt_schedule(&netisr_cpu
[i
]);
196 * The netisr_afree_rport is a special reply port which automatically
197 * frees the replied message. The netisr_adone_rport simply marks
198 * the message as being done. The netisr_apanic_rport panics if
199 * the message is replied to.
201 lwkt_initport_replyonly(&netisr_afree_rport
, netisr_autofree_reply
);
202 lwkt_initport_replyonly(&netisr_afree_free_so_rport
,
203 netisr_autofree_free_so_reply
);
204 lwkt_initport_replyonly_null(&netisr_adone_rport
);
205 lwkt_initport_panic(&netisr_apanic_rport
);
208 * The netisr_syncport is a special port which executes the message
209 * synchronously and waits for it if EASYNC is returned.
211 lwkt_initport_putonly(&netisr_sync_port
, netmsg_sync_putport
);
214 SYSINIT(netisr
, SI_SUB_PRE_DRIVERS
, SI_ORDER_FIRST
, netisr_init
, NULL
);
217 * Finish initializing the message port for a netmsg service. This also
218 * registers the port for synchronous cleanup operations such as when an
219 * ifnet is being destroyed. There is no deregistration API yet.
222 netmsg_service_port_init(lwkt_port_t port
)
224 struct netmsg_port_registration
*reg
;
227 * Override the putport function. Our custom function checks for
228 * self-references and executes such commands synchronously.
230 if (netmsg_fwd_port_fn
== NULL
)
231 netmsg_fwd_port_fn
= port
->mp_putport
;
232 KKASSERT(netmsg_fwd_port_fn
== port
->mp_putport
);
233 port
->mp_putport
= netmsg_put_port
;
236 * Keep track of ports using the netmsg API so we can synchronize
237 * certain operations (such as freeing an ifnet structure) across all
240 reg
= kmalloc(sizeof(*reg
), M_TEMP
, M_WAITOK
|M_ZERO
);
241 reg
->npr_port
= port
;
242 TAILQ_INSERT_TAIL(&netreglist
, reg
, npr_entry
);
246 * This function synchronizes the caller with all netmsg services. For
247 * example, if an interface is being removed we must make sure that all
248 * packets related to that interface complete processing before the structure
249 * can actually be freed. This sort of synchronization is an alternative to
250 * ref-counting the netif, removing the ref counting overhead in favor of
251 * placing additional overhead in the netif freeing sequence (where it is
255 netmsg_service_sync(void)
257 struct netmsg_port_registration
*reg
;
258 struct netmsg_base smsg
;
260 netmsg_init(&smsg
, NULL
, &curthread
->td_msgport
, 0, netmsg_sync_handler
);
262 TAILQ_FOREACH(reg
, &netreglist
, npr_entry
) {
263 lwkt_domsg(reg
->npr_port
, &smsg
.lmsg
, 0);
268 * The netmsg function simply replies the message. API semantics require
269 * EASYNC to be returned if the netmsg function disposes of the message.
272 netmsg_sync_handler(netmsg_t msg
)
274 lwkt_replymsg(&msg
->lmsg
, 0);
278 * Generic netmsg service loop. Some protocols may roll their own but all
279 * must do the basic command dispatch function call done here.
282 netmsg_service_loop(void *arg
)
284 struct netmsg_rollup
*ru
;
286 thread_t td
= curthread
;
289 td
->td_type
= TD_TYPE_NETISR
;
291 while ((msg
= lwkt_waitport(&td
->td_msgport
, 0))) {
293 * Run up to 512 pending netmsgs.
295 limit
= netisr_rollup_limit
;
297 KASSERT(msg
->nm_dispatch
!= NULL
,
298 ("netmsg_service isr %d badmsg",
299 msg
->lmsg
.u
.ms_result
));
301 * Don't match so_port, if the msg explicitly
302 * asks us to ignore its so_port.
304 if ((msg
->lmsg
.ms_flags
& MSGF_IGNSOPORT
) == 0 &&
306 msg
->nm_so
->so_port
!= &td
->td_msgport
) {
308 * Sockets undergoing connect or disconnect
309 * ops can change ports on us. Chase the
314 * This could be quite common for protocols
315 * which support asynchronous pru_connect,
316 * e.g. TCP, so kprintf socket port chasing
317 * could be too verbose for the console.
319 kprintf("%s: Warning, port changed so=%p\n",
320 __func__
, msg
->nm_so
);
322 lwkt_forwardmsg(msg
->nm_so
->so_port
,
326 * We are on the correct port, dispatch it.
328 msg
->nm_dispatch((netmsg_t
)msg
);
332 } while ((msg
= lwkt_getport(&td
->td_msgport
)) != NULL
);
335 * Run all registered rollup functions for this cpu
336 * (e.g. tcp_willblock()).
338 TAILQ_FOREACH(ru
, &netrulist
, ru_entry
)
344 * Forward a packet to a netisr service function.
346 * If the packet has not been assigned to a protocol thread we call
347 * the port characterization function to assign it. The caller must
348 * clear M_HASH (or not have set it in the first place) if the caller
349 * wishes the packet to be recharacterized.
352 netisr_queue(int num
, struct mbuf
*m
)
355 struct netmsg_packet
*pmsg
;
358 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
359 ("Bad isr %d", num
));
362 if (ni
->ni_handler
== NULL
) {
363 kprintf("%s: Unregistered isr %d\n", __func__
, num
);
369 * Figure out which protocol thread to send to. This does not
370 * have to be perfect but performance will be really good if it
371 * is correct. Major protocol inputs such as ip_input() will
372 * re-characterize the packet as necessary.
374 if ((m
->m_flags
& M_HASH
) == 0) {
375 ni
->ni_hashfn(&m
, 0);
378 if ((m
->m_flags
& M_HASH
) == 0) {
379 kprintf("%s(%d): packet hash failed\n",
387 * Get the protocol port based on the packet hash, initialize
388 * the netmsg, and send it off.
390 port
= netisr_hashport(m
->m_pkthdr
.hash
);
391 pmsg
= &m
->m_hdr
.mh_netmsg
;
392 netmsg_init(&pmsg
->base
, NULL
, &netisr_apanic_rport
,
395 pmsg
->base
.lmsg
.u
.ms_result
= num
;
396 lwkt_sendmsg(port
, &pmsg
->base
.lmsg
);
402 * Run a netisr service function on the packet.
404 * The packet must have been correctly characterized!
407 netisr_handle(int num
, struct mbuf
*m
)
410 struct netmsg_packet
*pmsg
;
414 * Get the protocol port based on the packet hash
416 KASSERT((m
->m_flags
& M_HASH
), ("packet not characterized"));
417 port
= netisr_hashport(m
->m_pkthdr
.hash
);
418 KASSERT(&curthread
->td_msgport
== port
, ("wrong msgport"));
420 KASSERT((num
> 0 && num
<= NELEM(netisrs
)), ("bad isr %d", num
));
422 if (ni
->ni_handler
== NULL
) {
423 kprintf("%s: unregistered isr %d\n", __func__
, num
);
429 * Initialize the netmsg, and run the handler directly.
431 pmsg
= &m
->m_hdr
.mh_netmsg
;
432 netmsg_init(&pmsg
->base
, NULL
, &netisr_apanic_rport
,
435 pmsg
->base
.lmsg
.u
.ms_result
= num
;
436 ni
->ni_handler((netmsg_t
)&pmsg
->base
);
442 * Pre-characterization of a deeper portion of the packet for the
445 * The base of the ISR type (e.g. IP) that we want to characterize is
446 * at (hoff) relative to the beginning of the mbuf. This allows
447 * e.g. ether_characterize() to not have to adjust the m_data/m_len.
450 netisr_characterize(int num
, struct mbuf
**mp
, int hoff
)
461 if (num
< 0 || num
>= NETISR_MAX
) {
462 if (num
== NETISR_MAX
) {
463 m
->m_flags
|= M_HASH
;
464 m
->m_pkthdr
.hash
= 0;
467 panic("Bad isr %d", num
);
474 if (ni
->ni_handler
== NULL
) {
475 kprintf("%s: Unregistered isr %d\n", __func__
, num
);
481 * Characterize the packet
483 if ((m
->m_flags
& M_HASH
) == 0) {
484 ni
->ni_hashfn(mp
, hoff
);
486 if (m
&& (m
->m_flags
& M_HASH
) == 0) {
487 kprintf("%s(%d): packet hash failed\n",
494 netisr_register(int num
, netisr_fn_t handler
, netisr_hashfn_t hashfn
)
498 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
499 ("netisr_register: bad isr %d", num
));
500 KKASSERT(handler
!= NULL
);
503 hashfn
= netisr_hashfn0
;
507 ni
->ni_handler
= handler
;
508 ni
->ni_hashck
= netisr_nohashck
;
509 ni
->ni_hashfn
= hashfn
;
510 netmsg_init(&ni
->ni_netmsg
, NULL
, &netisr_adone_rport
, 0, NULL
);
514 netisr_register_hashcheck(int num
, netisr_hashck_t hashck
)
518 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
519 ("netisr_register: bad isr %d", num
));
522 ni
->ni_hashck
= hashck
;
526 netisr_register_rollup(netisr_ru_t ru_func
, int prio
)
528 struct netmsg_rollup
*new_ru
, *ru
;
530 new_ru
= kmalloc(sizeof(*new_ru
), M_TEMP
, M_WAITOK
|M_ZERO
);
531 new_ru
->ru_func
= ru_func
;
532 new_ru
->ru_prio
= prio
;
535 * Higher priority "rollup" appears first
537 TAILQ_FOREACH(ru
, &netrulist
, ru_entry
) {
538 if (ru
->ru_prio
< new_ru
->ru_prio
) {
539 TAILQ_INSERT_BEFORE(ru
, new_ru
, ru_entry
);
543 TAILQ_INSERT_TAIL(&netrulist
, new_ru
, ru_entry
);
547 * Return a default protocol control message processing thread port
550 cpu0_ctlport(int cmd __unused
, struct sockaddr
*sa __unused
,
551 void *extra __unused
, int *cpuid
)
554 return netisr_cpuport(*cpuid
);
558 * This is a default netisr packet characterization function which
559 * sets M_HASH. If a netisr is registered with a NULL hashfn function
560 * this one is assigned.
562 * This function makes no attempt to validate the packet.
565 netisr_hashfn0(struct mbuf
**mp
, int hoff __unused
)
567 struct mbuf
*m
= *mp
;
569 m
->m_flags
|= M_HASH
;
570 m
->m_pkthdr
.hash
= 0;
574 * schednetisr() is used to call the netisr handler from the appropriate
575 * netisr thread for polling and other purposes.
577 * This function may be called from a hard interrupt or IPI and must be
578 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
579 * trying to allocate one. We must get ourselves onto the target cpu
580 * to safely check the MSGF_DONE bit on the message but since the message
581 * will be sent to that cpu anyway this does not add any extra work beyond
582 * what lwkt_sendmsg() would have already had to do to schedule the target
586 schednetisr_remote(void *data
)
588 int num
= (int)(intptr_t)data
;
589 struct netisr
*ni
= &netisrs
[num
];
590 lwkt_port_t port
= &netisr_cpu
[0].td_msgport
;
593 pmsg
= &netisrs
[num
].ni_netmsg
;
594 if (pmsg
->lmsg
.ms_flags
& MSGF_DONE
) {
595 netmsg_init(pmsg
, NULL
, &netisr_adone_rport
, 0, ni
->ni_handler
);
596 pmsg
->lmsg
.u
.ms_result
= num
;
597 lwkt_sendmsg(port
, &pmsg
->lmsg
);
604 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
605 ("schednetisr: bad isr %d", num
));
606 KKASSERT(netisrs
[num
].ni_handler
!= NULL
);
607 if (mycpu
->gd_cpuid
!= 0) {
608 lwkt_send_ipiq(globaldata_find(0),
609 schednetisr_remote
, (void *)(intptr_t)num
);
612 schednetisr_remote((void *)(intptr_t)num
);
618 netisr_barrier_dispatch(netmsg_t nmsg
)
620 struct netmsg_barrier
*msg
= (struct netmsg_barrier
*)nmsg
;
622 ATOMIC_CPUMASK_NANDBIT(*msg
->br_cpumask
, mycpu
->gd_cpuid
);
623 if (CPUMASK_TESTZERO(*msg
->br_cpumask
))
624 wakeup(msg
->br_cpumask
);
627 uint32_t done
= msg
->br_done
;
630 if ((done
& NETISR_BR_NOTDONE
) == 0)
633 tsleep_interlock(&msg
->br_done
, 0);
634 if (atomic_cmpset_int(&msg
->br_done
,
635 done
, done
| NETISR_BR_WAITDONE
))
636 tsleep(&msg
->br_done
, PINTERLOCKED
, "nbrdsp", 0);
639 lwkt_replymsg(&nmsg
->lmsg
, 0);
642 struct netisr_barrier
*
643 netisr_barrier_create(void)
645 struct netisr_barrier
*br
;
647 br
= kmalloc(sizeof(*br
), M_LWKTMSG
, M_WAITOK
| M_ZERO
);
652 netisr_barrier_set(struct netisr_barrier
*br
)
654 volatile cpumask_t other_cpumask
;
658 KKASSERT(!br
->br_isset
);
660 other_cpumask
= mycpu
->gd_other_cpus
;
661 CPUMASK_ANDMASK(other_cpumask
, smp_active_mask
);
664 for (i
= 0; i
< ncpus
; ++i
) {
665 struct netmsg_barrier
*msg
;
670 msg
= kmalloc(sizeof(struct netmsg_barrier
),
671 M_LWKTMSG
, M_WAITOK
);
674 * Don't use priority message here; mainly to keep
675 * it ordered w/ the previous data packets sent by
678 netmsg_init(&msg
->base
, NULL
, &netisr_afree_rport
, 0,
679 netisr_barrier_dispatch
);
680 msg
->br_cpumask
= &other_cpumask
;
681 msg
->br_done
= NETISR_BR_NOTDONE
;
683 KKASSERT(br
->br_msgs
[i
] == NULL
);
684 br
->br_msgs
[i
] = msg
;
687 for (i
= 0; i
< ncpus
; ++i
) {
690 lwkt_sendmsg(netisr_cpuport(i
), &br
->br_msgs
[i
]->base
.lmsg
);
693 while (CPUMASK_TESTNZERO(other_cpumask
)) {
694 tsleep_interlock(&other_cpumask
, 0);
695 if (CPUMASK_TESTNZERO(other_cpumask
))
696 tsleep(&other_cpumask
, PINTERLOCKED
, "nbrset", 0);
702 netisr_barrier_rem(struct netisr_barrier
*br
)
707 KKASSERT(br
->br_isset
);
710 for (i
= 0; i
< ncpus
; ++i
) {
711 struct netmsg_barrier
*msg
= br
->br_msgs
[i
];
714 msg
= br
->br_msgs
[i
];
715 br
->br_msgs
[i
] = NULL
;
720 done
= atomic_swap_int(&msg
->br_done
, 0);
721 if (done
& NETISR_BR_WAITDONE
)
722 wakeup(&msg
->br_done
);
728 netisr_nohashck(struct mbuf
*m
, const struct pktinfo
*pi __unused
)
730 m
->m_flags
&= ~M_HASH
;
734 netisr_hashcheck(int num
, struct mbuf
*m
, const struct pktinfo
*pi
)
738 if (num
< 0 || num
>= NETISR_MAX
)
739 panic("Bad isr %d", num
);
745 if (ni
->ni_handler
== NULL
)
746 panic("Unregistered isr %d", num
);
748 ni
->ni_hashck(m
, pi
);