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_port_init(lwkt_port_t
);
61 static void netmsg_service_loop(void *arg
);
62 static void netisr_hashfn0(struct mbuf
**mp
, int hoff
);
63 static void netisr_nohashck(struct mbuf
*, const struct pktinfo
*);
65 struct netmsg_port_registration
{
66 TAILQ_ENTRY(netmsg_port_registration
) npr_entry
;
70 struct netmsg_rollup
{
71 TAILQ_ENTRY(netmsg_rollup
) ru_entry
;
76 struct netmsg_barrier
{
77 struct netmsg_base base
;
78 volatile cpumask_t
*br_cpumask
;
79 volatile uint32_t br_done
;
82 #define NETISR_BR_NOTDONE 0x1
83 #define NETISR_BR_WAITDONE 0x80000000
85 struct netisr_barrier
{
86 struct netmsg_barrier
*br_msgs
[MAXCPU
];
90 void *netlastfunc
[MAXCPU
];
91 static struct netisr netisrs
[NETISR_MAX
];
92 static TAILQ_HEAD(,netmsg_port_registration
) netreglist
;
93 static TAILQ_HEAD(,netmsg_rollup
) netrulist
;
95 /* Per-CPU thread to handle any protocol. */
96 struct thread
*netisr_threads
[MAXCPU
];
98 lwkt_port netisr_afree_rport
;
99 lwkt_port netisr_afree_free_so_rport
;
100 lwkt_port netisr_adone_rport
;
101 lwkt_port netisr_apanic_rport
;
102 lwkt_port netisr_sync_port
;
104 static int (*netmsg_fwd_port_fn
)(lwkt_port_t
, lwkt_msg_t
);
106 SYSCTL_NODE(_net
, OID_AUTO
, netisr
, CTLFLAG_RW
, 0, "netisr");
108 static int netisr_rollup_limit
= 32;
109 SYSCTL_INT(_net_netisr
, OID_AUTO
, rollup_limit
, CTLFLAG_RW
,
110 &netisr_rollup_limit
, 0, "Message to process before rollup");
113 TUNABLE_INT("net.netisr.ncpus", &netisr_ncpus
);
114 SYSCTL_INT(_net_netisr
, OID_AUTO
, ncpus
, CTLFLAG_RD
,
115 &netisr_ncpus
, 0, "# of CPUs to handle network messages");
118 * netisr_afree_rport replymsg function, only used to handle async
119 * messages which the sender has abandoned to their fate.
122 netisr_autofree_reply(lwkt_port_t port
, lwkt_msg_t msg
)
124 kfree(msg
, M_LWKTMSG
);
128 netisr_autofree_free_so_reply(lwkt_port_t port
, lwkt_msg_t msg
)
130 sofree(((netmsg_t
)msg
)->base
.nm_so
);
131 kfree(msg
, M_LWKTMSG
);
135 * We need a custom putport function to handle the case where the
136 * message target is the current thread's message port. This case
137 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
138 * then turns around and executes a network operation synchronously.
140 * To prevent deadlocking, we must execute these self-referential messages
141 * synchronously, effectively turning the message into a glorified direct
142 * procedure call back into the protocol stack. The operation must be
143 * complete on return or we will deadlock, so panic if it isn't.
145 * However, the target function is under no obligation to immediately
146 * reply the message. It may forward it elsewhere.
149 netmsg_put_port(lwkt_port_t port
, lwkt_msg_t lmsg
)
151 netmsg_base_t nmsg
= (void *)lmsg
;
153 if ((lmsg
->ms_flags
& MSGF_SYNC
) && port
== &curthread
->td_msgport
) {
154 nmsg
->nm_dispatch((netmsg_t
)nmsg
);
157 return(netmsg_fwd_port_fn(port
, lmsg
));
162 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
163 * because they depend on the user proc context for a number of things
164 * (like creds) which we have not yet incorporated into the message structure.
166 * However, we maintain or message/port abstraction. Having a special
167 * synchronous port which runs the commands synchronously gives us the
168 * ability to serialize operations in one place later on when we start
172 netmsg_sync_putport(lwkt_port_t port
, lwkt_msg_t lmsg
)
174 netmsg_base_t nmsg
= (void *)lmsg
;
176 KKASSERT((lmsg
->ms_flags
& MSGF_DONE
) == 0);
178 lmsg
->ms_target_port
= port
; /* required for abort */
179 nmsg
->nm_dispatch((netmsg_t
)nmsg
);
188 if (netisr_ncpus
<= 0) {
190 netisr_ncpus
= ncpus2
;
191 } else if (netisr_ncpus
> ncpus
) {
192 netisr_ncpus
= ncpus
;
194 if (netisr_ncpus
> NETISR_CPUMAX
)
195 netisr_ncpus
= NETISR_CPUMAX
;
197 TAILQ_INIT(&netreglist
);
198 TAILQ_INIT(&netrulist
);
201 * Create default per-cpu threads for generic protocol handling.
203 for (i
= 0; i
< ncpus
; ++i
) {
204 lwkt_create(netmsg_service_loop
, NULL
, &netisr_threads
[i
], NULL
,
205 TDF_NOSTART
|TDF_FORCE_SPINPORT
|TDF_FIXEDCPU
,
207 netmsg_service_port_init(&netisr_threads
[i
]->td_msgport
);
208 lwkt_schedule(netisr_threads
[i
]);
212 * The netisr_afree_rport is a special reply port which automatically
213 * frees the replied message. The netisr_adone_rport simply marks
214 * the message as being done. The netisr_apanic_rport panics if
215 * the message is replied to.
217 lwkt_initport_replyonly(&netisr_afree_rport
, netisr_autofree_reply
);
218 lwkt_initport_replyonly(&netisr_afree_free_so_rport
,
219 netisr_autofree_free_so_reply
);
220 lwkt_initport_replyonly_null(&netisr_adone_rport
);
221 lwkt_initport_panic(&netisr_apanic_rport
);
224 * The netisr_syncport is a special port which executes the message
225 * synchronously and waits for it if EASYNC is returned.
227 lwkt_initport_putonly(&netisr_sync_port
, netmsg_sync_putport
);
229 SYSINIT(netisr
, SI_SUB_PRE_DRIVERS
, SI_ORDER_FIRST
, netisr_init
, NULL
);
232 * Finish initializing the message port for a netmsg service. This also
233 * registers the port for synchronous cleanup operations such as when an
234 * ifnet is being destroyed. There is no deregistration API yet.
237 netmsg_service_port_init(lwkt_port_t port
)
239 struct netmsg_port_registration
*reg
;
242 * Override the putport function. Our custom function checks for
243 * self-references and executes such commands synchronously.
245 if (netmsg_fwd_port_fn
== NULL
)
246 netmsg_fwd_port_fn
= port
->mp_putport
;
247 KKASSERT(netmsg_fwd_port_fn
== port
->mp_putport
);
248 port
->mp_putport
= netmsg_put_port
;
251 * Keep track of ports using the netmsg API so we can synchronize
252 * certain operations (such as freeing an ifnet structure) across all
255 reg
= kmalloc(sizeof(*reg
), M_TEMP
, M_WAITOK
|M_ZERO
);
256 reg
->npr_port
= port
;
257 TAILQ_INSERT_TAIL(&netreglist
, reg
, npr_entry
);
261 * This function synchronizes the caller with all netmsg services. For
262 * example, if an interface is being removed we must make sure that all
263 * packets related to that interface complete processing before the structure
264 * can actually be freed. This sort of synchronization is an alternative to
265 * ref-counting the netif, removing the ref counting overhead in favor of
266 * placing additional overhead in the netif freeing sequence (where it is
270 netmsg_service_sync(void)
272 struct netmsg_port_registration
*reg
;
273 struct netmsg_base smsg
;
275 netmsg_init(&smsg
, NULL
, &curthread
->td_msgport
, 0, netmsg_sync_handler
);
277 TAILQ_FOREACH(reg
, &netreglist
, npr_entry
) {
278 lwkt_domsg(reg
->npr_port
, &smsg
.lmsg
, 0);
283 * The netmsg function simply replies the message. API semantics require
284 * EASYNC to be returned if the netmsg function disposes of the message.
287 netmsg_sync_handler(netmsg_t msg
)
289 lwkt_replymsg(&msg
->lmsg
, 0);
293 * Generic netmsg service loop. Some protocols may roll their own but all
294 * must do the basic command dispatch function call done here.
297 netmsg_service_loop(void *arg
)
299 struct netmsg_rollup
*ru
;
301 thread_t td
= curthread
;
304 td
->td_type
= TD_TYPE_NETISR
;
306 while ((msg
= lwkt_waitport(&td
->td_msgport
, 0))) {
308 * Run up to 512 pending netmsgs.
310 limit
= netisr_rollup_limit
;
312 KASSERT(msg
->nm_dispatch
!= NULL
,
313 ("netmsg_service isr %d badmsg",
314 msg
->lmsg
.u
.ms_result
));
316 * Don't match so_port, if the msg explicitly
317 * asks us to ignore its so_port.
319 if ((msg
->lmsg
.ms_flags
& MSGF_IGNSOPORT
) == 0 &&
321 msg
->nm_so
->so_port
!= &td
->td_msgport
) {
323 * Sockets undergoing connect or disconnect
324 * ops can change ports on us. Chase the
329 * This could be quite common for protocols
330 * which support asynchronous pru_connect,
331 * e.g. TCP, so kprintf socket port chasing
332 * could be too verbose for the console.
334 kprintf("%s: Warning, port changed so=%p\n",
335 __func__
, msg
->nm_so
);
337 lwkt_forwardmsg(msg
->nm_so
->so_port
,
341 * We are on the correct port, dispatch it.
343 netlastfunc
[mycpuid
] = msg
->nm_dispatch
;
344 msg
->nm_dispatch((netmsg_t
)msg
);
348 } while ((msg
= lwkt_getport(&td
->td_msgport
)) != NULL
);
351 * Run all registered rollup functions for this cpu
352 * (e.g. tcp_willblock()).
354 TAILQ_FOREACH(ru
, &netrulist
, ru_entry
)
360 * Forward a packet to a netisr service function.
362 * If the packet has not been assigned to a protocol thread we call
363 * the port characterization function to assign it. The caller must
364 * clear M_HASH (or not have set it in the first place) if the caller
365 * wishes the packet to be recharacterized.
368 netisr_queue(int num
, struct mbuf
*m
)
371 struct netmsg_packet
*pmsg
;
374 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
375 ("Bad isr %d", num
));
378 if (ni
->ni_handler
== NULL
) {
379 kprintf("%s: Unregistered isr %d\n", __func__
, num
);
385 * Figure out which protocol thread to send to. This does not
386 * have to be perfect but performance will be really good if it
387 * is correct. Major protocol inputs such as ip_input() will
388 * re-characterize the packet as necessary.
390 if ((m
->m_flags
& M_HASH
) == 0) {
391 ni
->ni_hashfn(&m
, 0);
394 if ((m
->m_flags
& M_HASH
) == 0) {
395 kprintf("%s(%d): packet hash failed\n",
403 * Get the protocol port based on the packet hash, initialize
404 * the netmsg, and send it off.
406 port
= netisr_hashport(m
->m_pkthdr
.hash
);
407 pmsg
= &m
->m_hdr
.mh_netmsg
;
408 netmsg_init(&pmsg
->base
, NULL
, &netisr_apanic_rport
,
411 pmsg
->base
.lmsg
.u
.ms_result
= num
;
412 lwkt_sendmsg(port
, &pmsg
->base
.lmsg
);
418 * Run a netisr service function on the packet.
420 * The packet must have been correctly characterized!
423 netisr_handle(int num
, struct mbuf
*m
)
426 struct netmsg_packet
*pmsg
;
430 * Get the protocol port based on the packet hash
432 KASSERT((m
->m_flags
& M_HASH
), ("packet not characterized"));
433 port
= netisr_hashport(m
->m_pkthdr
.hash
);
434 KASSERT(&curthread
->td_msgport
== port
, ("wrong msgport"));
436 KASSERT((num
> 0 && num
<= NELEM(netisrs
)), ("bad isr %d", num
));
438 if (ni
->ni_handler
== NULL
) {
439 kprintf("%s: unregistered isr %d\n", __func__
, num
);
445 * Initialize the netmsg, and run the handler directly.
447 pmsg
= &m
->m_hdr
.mh_netmsg
;
448 netmsg_init(&pmsg
->base
, NULL
, &netisr_apanic_rport
,
451 pmsg
->base
.lmsg
.u
.ms_result
= num
;
452 ni
->ni_handler((netmsg_t
)&pmsg
->base
);
458 * Pre-characterization of a deeper portion of the packet for the
461 * The base of the ISR type (e.g. IP) that we want to characterize is
462 * at (hoff) relative to the beginning of the mbuf. This allows
463 * e.g. ether_characterize() to not have to adjust the m_data/m_len.
466 netisr_characterize(int num
, struct mbuf
**mp
, int hoff
)
477 if (num
< 0 || num
>= NETISR_MAX
) {
478 if (num
== NETISR_MAX
) {
482 panic("Bad isr %d", num
);
489 if (ni
->ni_handler
== NULL
) {
490 kprintf("%s: Unregistered isr %d\n", __func__
, num
);
496 * Characterize the packet
498 if ((m
->m_flags
& M_HASH
) == 0) {
499 ni
->ni_hashfn(mp
, hoff
);
501 if (m
&& (m
->m_flags
& M_HASH
) == 0) {
502 kprintf("%s(%d): packet hash failed\n",
509 netisr_register(int num
, netisr_fn_t handler
, netisr_hashfn_t hashfn
)
513 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
514 ("netisr_register: bad isr %d", num
));
515 KKASSERT(handler
!= NULL
);
518 hashfn
= netisr_hashfn0
;
522 ni
->ni_handler
= handler
;
523 ni
->ni_hashck
= netisr_nohashck
;
524 ni
->ni_hashfn
= hashfn
;
525 netmsg_init(&ni
->ni_netmsg
, NULL
, &netisr_adone_rport
, 0, NULL
);
529 netisr_register_hashcheck(int num
, netisr_hashck_t hashck
)
533 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
534 ("netisr_register: bad isr %d", num
));
537 ni
->ni_hashck
= hashck
;
541 netisr_register_rollup(netisr_ru_t ru_func
, int prio
)
543 struct netmsg_rollup
*new_ru
, *ru
;
545 new_ru
= kmalloc(sizeof(*new_ru
), M_TEMP
, M_WAITOK
|M_ZERO
);
546 new_ru
->ru_func
= ru_func
;
547 new_ru
->ru_prio
= prio
;
550 * Higher priority "rollup" appears first
552 TAILQ_FOREACH(ru
, &netrulist
, ru_entry
) {
553 if (ru
->ru_prio
< new_ru
->ru_prio
) {
554 TAILQ_INSERT_BEFORE(ru
, new_ru
, ru_entry
);
558 TAILQ_INSERT_TAIL(&netrulist
, new_ru
, ru_entry
);
562 * Return a default protocol control message processing thread port
565 cpu0_ctlport(int cmd __unused
, struct sockaddr
*sa __unused
,
566 void *extra __unused
, int *cpuid
)
569 return netisr_cpuport(*cpuid
);
573 * This is a default netisr packet characterization function which
574 * sets M_HASH. If a netisr is registered with a NULL hashfn function
575 * this one is assigned.
577 * This function makes no attempt to validate the packet.
580 netisr_hashfn0(struct mbuf
**mp
, int hoff __unused
)
587 * schednetisr() is used to call the netisr handler from the appropriate
588 * netisr thread for polling and other purposes.
590 * This function may be called from a hard interrupt or IPI and must be
591 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
592 * trying to allocate one. We must get ourselves onto the target cpu
593 * to safely check the MSGF_DONE bit on the message but since the message
594 * will be sent to that cpu anyway this does not add any extra work beyond
595 * what lwkt_sendmsg() would have already had to do to schedule the target
599 schednetisr_remote(void *data
)
601 int num
= (int)(intptr_t)data
;
602 struct netisr
*ni
= &netisrs
[num
];
603 lwkt_port_t port
= &netisr_threads
[0]->td_msgport
;
606 pmsg
= &netisrs
[num
].ni_netmsg
;
607 if (pmsg
->lmsg
.ms_flags
& MSGF_DONE
) {
608 netmsg_init(pmsg
, NULL
, &netisr_adone_rport
, 0, ni
->ni_handler
);
609 pmsg
->lmsg
.u
.ms_result
= num
;
610 lwkt_sendmsg(port
, &pmsg
->lmsg
);
617 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
618 ("schednetisr: bad isr %d", num
));
619 KKASSERT(netisrs
[num
].ni_handler
!= NULL
);
620 if (mycpu
->gd_cpuid
!= 0) {
621 lwkt_send_ipiq(globaldata_find(0),
622 schednetisr_remote
, (void *)(intptr_t)num
);
625 schednetisr_remote((void *)(intptr_t)num
);
631 netisr_barrier_dispatch(netmsg_t nmsg
)
633 struct netmsg_barrier
*msg
= (struct netmsg_barrier
*)nmsg
;
635 ATOMIC_CPUMASK_NANDBIT(*msg
->br_cpumask
, mycpu
->gd_cpuid
);
636 if (CPUMASK_TESTZERO(*msg
->br_cpumask
))
637 wakeup(msg
->br_cpumask
);
640 uint32_t done
= msg
->br_done
;
643 if ((done
& NETISR_BR_NOTDONE
) == 0)
646 tsleep_interlock(&msg
->br_done
, 0);
647 if (atomic_cmpset_int(&msg
->br_done
,
648 done
, done
| NETISR_BR_WAITDONE
))
649 tsleep(&msg
->br_done
, PINTERLOCKED
, "nbrdsp", 0);
652 lwkt_replymsg(&nmsg
->lmsg
, 0);
655 struct netisr_barrier
*
656 netisr_barrier_create(void)
658 struct netisr_barrier
*br
;
660 br
= kmalloc(sizeof(*br
), M_LWKTMSG
, M_WAITOK
| M_ZERO
);
665 netisr_barrier_set(struct netisr_barrier
*br
)
667 volatile cpumask_t other_cpumask
;
671 KKASSERT(!br
->br_isset
);
673 other_cpumask
= mycpu
->gd_other_cpus
;
674 CPUMASK_ANDMASK(other_cpumask
, smp_active_mask
);
677 for (i
= 0; i
< ncpus
; ++i
) {
678 struct netmsg_barrier
*msg
;
683 msg
= kmalloc(sizeof(struct netmsg_barrier
),
684 M_LWKTMSG
, M_WAITOK
);
687 * Don't use priority message here; mainly to keep
688 * it ordered w/ the previous data packets sent by
691 netmsg_init(&msg
->base
, NULL
, &netisr_afree_rport
, 0,
692 netisr_barrier_dispatch
);
693 msg
->br_cpumask
= &other_cpumask
;
694 msg
->br_done
= NETISR_BR_NOTDONE
;
696 KKASSERT(br
->br_msgs
[i
] == NULL
);
697 br
->br_msgs
[i
] = msg
;
700 for (i
= 0; i
< ncpus
; ++i
) {
703 lwkt_sendmsg(netisr_cpuport(i
), &br
->br_msgs
[i
]->base
.lmsg
);
706 while (CPUMASK_TESTNZERO(other_cpumask
)) {
707 tsleep_interlock(&other_cpumask
, 0);
708 if (CPUMASK_TESTNZERO(other_cpumask
))
709 tsleep(&other_cpumask
, PINTERLOCKED
, "nbrset", 0);
715 netisr_barrier_rem(struct netisr_barrier
*br
)
720 KKASSERT(br
->br_isset
);
723 for (i
= 0; i
< ncpus
; ++i
) {
724 struct netmsg_barrier
*msg
= br
->br_msgs
[i
];
727 msg
= br
->br_msgs
[i
];
728 br
->br_msgs
[i
] = NULL
;
733 done
= atomic_swap_int(&msg
->br_done
, 0);
734 if (done
& NETISR_BR_WAITDONE
)
735 wakeup(&msg
->br_done
);
741 netisr_nohashck(struct mbuf
*m
, const struct pktinfo
*pi __unused
)
743 m
->m_flags
&= ~M_HASH
;
747 netisr_hashcheck(int num
, struct mbuf
*m
, const struct pktinfo
*pi
)
751 if (num
< 0 || num
>= NETISR_MAX
)
752 panic("Bad isr %d", num
);
758 if (ni
->ni_handler
== NULL
)
759 panic("Unregistered isr %d", num
);
761 ni
->ni_hashck(m
, pi
);