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/netisr.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 cpu0_cpufn(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
;
74 struct netmsg_barrier
{
75 struct netmsg_base base
;
76 volatile cpumask_t
*br_cpumask
;
77 volatile uint32_t br_done
;
80 #define NETISR_BR_NOTDONE 0x1
81 #define NETISR_BR_WAITDONE 0x80000000
83 struct netisr_barrier
{
84 struct netmsg_barrier
*br_msgs
[MAXCPU
];
88 static struct netisr netisrs
[NETISR_MAX
];
89 static TAILQ_HEAD(,netmsg_port_registration
) netreglist
;
90 static TAILQ_HEAD(,netmsg_rollup
) netrulist
;
92 /* Per-CPU thread to handle any protocol. */
93 static struct thread netisr_cpu
[MAXCPU
];
94 lwkt_port netisr_afree_rport
;
95 lwkt_port netisr_afree_free_so_rport
;
96 lwkt_port netisr_adone_rport
;
97 lwkt_port netisr_apanic_rport
;
98 lwkt_port netisr_sync_port
;
100 static int (*netmsg_fwd_port_fn
)(lwkt_port_t
, lwkt_msg_t
);
102 SYSCTL_NODE(_net
, OID_AUTO
, netisr
, CTLFLAG_RW
, 0, "netisr");
105 * netisr_afree_rport replymsg function, only used to handle async
106 * messages which the sender has abandoned to their fate.
109 netisr_autofree_reply(lwkt_port_t port
, lwkt_msg_t msg
)
111 kfree(msg
, M_LWKTMSG
);
115 netisr_autofree_free_so_reply(lwkt_port_t port
, lwkt_msg_t msg
)
117 sofree(((netmsg_t
)msg
)->base
.nm_so
);
118 kfree(msg
, M_LWKTMSG
);
122 * We need a custom putport function to handle the case where the
123 * message target is the current thread's message port. This case
124 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
125 * then turns around and executes a network operation synchronously.
127 * To prevent deadlocking, we must execute these self-referential messages
128 * synchronously, effectively turning the message into a glorified direct
129 * procedure call back into the protocol stack. The operation must be
130 * complete on return or we will deadlock, so panic if it isn't.
132 * However, the target function is under no obligation to immediately
133 * reply the message. It may forward it elsewhere.
136 netmsg_put_port(lwkt_port_t port
, lwkt_msg_t lmsg
)
138 netmsg_base_t nmsg
= (void *)lmsg
;
140 if ((lmsg
->ms_flags
& MSGF_SYNC
) && port
== &curthread
->td_msgport
) {
141 nmsg
->nm_dispatch((netmsg_t
)nmsg
);
144 return(netmsg_fwd_port_fn(port
, lmsg
));
149 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
150 * because they depend on the user proc context for a number of things
151 * (like creds) which we have not yet incorporated into the message structure.
153 * However, we maintain or message/port abstraction. Having a special
154 * synchronous port which runs the commands synchronously gives us the
155 * ability to serialize operations in one place later on when we start
159 netmsg_sync_putport(lwkt_port_t port
, lwkt_msg_t lmsg
)
161 netmsg_base_t nmsg
= (void *)lmsg
;
163 KKASSERT((lmsg
->ms_flags
& MSGF_DONE
) == 0);
165 lmsg
->ms_target_port
= port
; /* required for abort */
166 nmsg
->nm_dispatch((netmsg_t
)nmsg
);
175 TAILQ_INIT(&netreglist
);
176 TAILQ_INIT(&netrulist
);
179 * Create default per-cpu threads for generic protocol handling.
181 for (i
= 0; i
< ncpus
; ++i
) {
182 lwkt_create(netmsg_service_loop
, NULL
, NULL
,
183 &netisr_cpu
[i
], TDF_NOSTART
|TDF_FORCE_SPINPORT
,
184 i
, "netisr_cpu %d", i
);
185 netmsg_service_port_init(&netisr_cpu
[i
].td_msgport
);
186 lwkt_schedule(&netisr_cpu
[i
]);
190 * The netisr_afree_rport is a special reply port which automatically
191 * frees the replied message. The netisr_adone_rport simply marks
192 * the message as being done. The netisr_apanic_rport panics if
193 * the message is replied to.
195 lwkt_initport_replyonly(&netisr_afree_rport
, netisr_autofree_reply
);
196 lwkt_initport_replyonly(&netisr_afree_free_so_rport
,
197 netisr_autofree_free_so_reply
);
198 lwkt_initport_replyonly_null(&netisr_adone_rport
);
199 lwkt_initport_panic(&netisr_apanic_rport
);
202 * The netisr_syncport is a special port which executes the message
203 * synchronously and waits for it if EASYNC is returned.
205 lwkt_initport_putonly(&netisr_sync_port
, netmsg_sync_putport
);
208 SYSINIT(netisr
, SI_SUB_PRE_DRIVERS
, SI_ORDER_FIRST
, netisr_init
, NULL
);
211 * Finish initializing the message port for a netmsg service. This also
212 * registers the port for synchronous cleanup operations such as when an
213 * ifnet is being destroyed. There is no deregistration API yet.
216 netmsg_service_port_init(lwkt_port_t port
)
218 struct netmsg_port_registration
*reg
;
221 * Override the putport function. Our custom function checks for
222 * self-references and executes such commands synchronously.
224 if (netmsg_fwd_port_fn
== NULL
)
225 netmsg_fwd_port_fn
= port
->mp_putport
;
226 KKASSERT(netmsg_fwd_port_fn
== port
->mp_putport
);
227 port
->mp_putport
= netmsg_put_port
;
230 * Keep track of ports using the netmsg API so we can synchronize
231 * certain operations (such as freeing an ifnet structure) across all
234 reg
= kmalloc(sizeof(*reg
), M_TEMP
, M_WAITOK
|M_ZERO
);
235 reg
->npr_port
= port
;
236 TAILQ_INSERT_TAIL(&netreglist
, reg
, npr_entry
);
240 * This function synchronizes the caller with all netmsg services. For
241 * example, if an interface is being removed we must make sure that all
242 * packets related to that interface complete processing before the structure
243 * can actually be freed. This sort of synchronization is an alternative to
244 * ref-counting the netif, removing the ref counting overhead in favor of
245 * placing additional overhead in the netif freeing sequence (where it is
249 netmsg_service_sync(void)
251 struct netmsg_port_registration
*reg
;
252 struct netmsg_base smsg
;
254 netmsg_init(&smsg
, NULL
, &curthread
->td_msgport
, 0, netmsg_sync_handler
);
256 TAILQ_FOREACH(reg
, &netreglist
, npr_entry
) {
257 lwkt_domsg(reg
->npr_port
, &smsg
.lmsg
, 0);
262 * The netmsg function simply replies the message. API semantics require
263 * EASYNC to be returned if the netmsg function disposes of the message.
266 netmsg_sync_handler(netmsg_t msg
)
268 lwkt_replymsg(&msg
->lmsg
, 0);
272 * Generic netmsg service loop. Some protocols may roll their own but all
273 * must do the basic command dispatch function call done here.
276 netmsg_service_loop(void *arg
)
278 struct netmsg_rollup
*ru
;
280 thread_t td
= curthread
;;
283 while ((msg
= lwkt_waitport(&td
->td_msgport
, 0))) {
285 * Run up to 512 pending netmsgs.
289 KASSERT(msg
->nm_dispatch
!= NULL
,
290 ("netmsg_service isr %d badmsg",
291 msg
->lmsg
.u
.ms_result
));
293 msg
->nm_so
->so_port
!= &td
->td_msgport
) {
295 * Sockets undergoing connect or disconnect
296 * ops can change ports on us. Chase the
299 kprintf("netmsg_service_loop: Warning, "
300 "port changed so=%p\n", msg
->nm_so
);
301 lwkt_forwardmsg(msg
->nm_so
->so_port
,
305 * We are on the correct port, dispatch it.
307 msg
->nm_dispatch((netmsg_t
)msg
);
311 } while ((msg
= lwkt_getport(&td
->td_msgport
)) != NULL
);
314 * Run all registered rollup functions for this cpu
315 * (e.g. tcp_willblock()).
317 TAILQ_FOREACH(ru
, &netrulist
, ru_entry
)
323 * Forward a packet to a netisr service function.
325 * If the packet has not been assigned to a protocol thread we call
326 * the port characterization function to assign it. The caller must
327 * clear M_HASH (or not have set it in the first place) if the caller
328 * wishes the packet to be recharacterized.
331 netisr_queue(int num
, struct mbuf
*m
)
334 struct netmsg_packet
*pmsg
;
337 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
338 ("Bad isr %d", num
));
341 if (ni
->ni_handler
== NULL
) {
342 kprintf("Unregistered isr %d\n", num
);
348 * Figure out which protocol thread to send to. This does not
349 * have to be perfect but performance will be really good if it
350 * is correct. Major protocol inputs such as ip_input() will
351 * re-characterize the packet as necessary.
353 if ((m
->m_flags
& M_HASH
) == 0) {
359 if ((m
->m_flags
& M_HASH
) == 0) {
360 kprintf("netisr_queue(%d): packet hash failed\n", num
);
367 * Get the protocol port based on the packet hash, initialize
368 * the netmsg, and send it off.
370 port
= netisr_portfn(m
->m_pkthdr
.hash
);
371 pmsg
= &m
->m_hdr
.mh_netmsg
;
372 netmsg_init(&pmsg
->base
, NULL
, &netisr_apanic_rport
,
375 pmsg
->base
.lmsg
.u
.ms_result
= num
;
376 lwkt_sendmsg(port
, &pmsg
->base
.lmsg
);
382 * Run a netisr service function on the packet.
384 * The packet must have been correctly characterized!
387 netisr_handle(int num
, struct mbuf
*m
)
390 struct netmsg_packet
*pmsg
;
394 * Get the protocol port based on the packet hash
396 KASSERT((m
->m_flags
& M_HASH
), ("packet not characterized"));
397 port
= netisr_portfn(m
->m_pkthdr
.hash
);
398 KASSERT(&curthread
->td_msgport
== port
, ("wrong msgport"));
400 KASSERT((num
> 0 && num
<= NELEM(netisrs
)), ("bad isr %d", num
));
402 if (ni
->ni_handler
== NULL
) {
403 kprintf("unregistered isr %d\n", num
);
409 * Initialize the netmsg, and run the handler directly.
411 pmsg
= &m
->m_hdr
.mh_netmsg
;
412 netmsg_init(&pmsg
->base
, NULL
, &netisr_apanic_rport
,
415 pmsg
->base
.lmsg
.u
.ms_result
= num
;
416 ni
->ni_handler((netmsg_t
)&pmsg
->base
);
422 * Pre-characterization of a deeper portion of the packet for the
425 * The base of the ISR type (e.g. IP) that we want to characterize is
426 * at (hoff) relative to the beginning of the mbuf. This allows
427 * e.g. ether_input_chain() to not have to adjust the m_data/m_len.
430 netisr_characterize(int num
, struct mbuf
**mp
, int hoff
)
441 if (num
< 0 || num
>= NETISR_MAX
) {
442 if (num
== NETISR_MAX
) {
443 m
->m_flags
|= M_HASH
;
444 m
->m_pkthdr
.hash
= 0;
447 panic("Bad isr %d", num
);
454 if (ni
->ni_handler
== NULL
) {
455 kprintf("Unregistered isr %d\n", num
);
461 * Characterize the packet
463 if ((m
->m_flags
& M_HASH
) == 0) {
464 ni
->ni_cpufn(mp
, hoff
);
466 if (m
&& (m
->m_flags
& M_HASH
) == 0)
467 kprintf("netisr_queue(%d): packet hash failed\n", num
);
472 netisr_register(int num
, netisr_fn_t handler
, netisr_cpufn_t cpufn
)
476 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
477 ("netisr_register: bad isr %d", num
));
478 KKASSERT(handler
!= NULL
);
485 ni
->ni_handler
= handler
;
486 ni
->ni_hashck
= netisr_nohashck
;
487 ni
->ni_cpufn
= cpufn
;
488 netmsg_init(&ni
->ni_netmsg
, NULL
, &netisr_adone_rport
, 0, NULL
);
492 netisr_register_hashcheck(int num
, netisr_hashck_t hashck
)
496 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
497 ("netisr_register: bad isr %d", num
));
500 ni
->ni_hashck
= hashck
;
504 netisr_register_rollup(netisr_ru_t ru_func
)
506 struct netmsg_rollup
*ru
;
508 ru
= kmalloc(sizeof(*ru
), M_TEMP
, M_WAITOK
|M_ZERO
);
509 ru
->ru_func
= ru_func
;
510 TAILQ_INSERT_TAIL(&netrulist
, ru
, ru_entry
);
514 * Return the message port for the general protocol message servicing
515 * thread for a particular cpu.
518 netisr_portfn(int cpu
)
520 KKASSERT(cpu
>= 0 && cpu
< ncpus
);
521 return (&netisr_cpu
[cpu
].td_msgport
);
525 * Return the current cpu's network protocol thread.
530 return(netisr_portfn(mycpu
->gd_cpuid
));
534 * Return a default protocol control message processing thread port
537 cpu0_ctlport(int cmd __unused
, struct sockaddr
*sa __unused
,
538 void *extra __unused
)
540 return (&netisr_cpu
[0].td_msgport
);
544 * This is a default netisr packet characterization function which
545 * sets M_HASH. If a netisr is registered with a NULL cpufn function
546 * this one is assigned.
548 * This function makes no attempt to validate the packet.
551 cpu0_cpufn(struct mbuf
**mp
, int hoff __unused
)
553 struct mbuf
*m
= *mp
;
555 m
->m_flags
|= M_HASH
;
556 m
->m_pkthdr
.hash
= 0;
560 * schednetisr() is used to call the netisr handler from the appropriate
561 * netisr thread for polling and other purposes.
563 * This function may be called from a hard interrupt or IPI and must be
564 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
565 * trying to allocate one. We must get ourselves onto the target cpu
566 * to safely check the MSGF_DONE bit on the message but since the message
567 * will be sent to that cpu anyway this does not add any extra work beyond
568 * what lwkt_sendmsg() would have already had to do to schedule the target
572 schednetisr_remote(void *data
)
574 int num
= (int)(intptr_t)data
;
575 struct netisr
*ni
= &netisrs
[num
];
576 lwkt_port_t port
= &netisr_cpu
[0].td_msgport
;
579 pmsg
= &netisrs
[num
].ni_netmsg
;
580 if (pmsg
->lmsg
.ms_flags
& MSGF_DONE
) {
581 netmsg_init(pmsg
, NULL
, &netisr_adone_rport
, 0, ni
->ni_handler
);
582 pmsg
->lmsg
.u
.ms_result
= num
;
583 lwkt_sendmsg(port
, &pmsg
->lmsg
);
590 KASSERT((num
> 0 && num
<= NELEM(netisrs
)),
591 ("schednetisr: bad isr %d", num
));
592 KKASSERT(netisrs
[num
].ni_handler
!= NULL
);
594 if (mycpu
->gd_cpuid
!= 0) {
595 lwkt_send_ipiq(globaldata_find(0),
596 schednetisr_remote
, (void *)(intptr_t)num
);
599 schednetisr_remote((void *)(intptr_t)num
);
604 schednetisr_remote((void *)(intptr_t)num
);
612 netisr_barrier_dispatch(netmsg_t nmsg
)
614 struct netmsg_barrier
*msg
= (struct netmsg_barrier
*)nmsg
;
616 atomic_clear_cpumask(msg
->br_cpumask
, mycpu
->gd_cpumask
);
617 if (*msg
->br_cpumask
== 0)
618 wakeup(msg
->br_cpumask
);
621 uint32_t done
= msg
->br_done
;
624 if ((done
& NETISR_BR_NOTDONE
) == 0)
627 tsleep_interlock(&msg
->br_done
, 0);
628 if (atomic_cmpset_int(&msg
->br_done
,
629 done
, done
| NETISR_BR_WAITDONE
))
630 tsleep(&msg
->br_done
, PINTERLOCKED
, "nbrdsp", 0);
633 lwkt_replymsg(&nmsg
->lmsg
, 0);
638 struct netisr_barrier
*
639 netisr_barrier_create(void)
641 struct netisr_barrier
*br
;
643 br
= kmalloc(sizeof(*br
), M_LWKTMSG
, M_WAITOK
| M_ZERO
);
648 netisr_barrier_set(struct netisr_barrier
*br
)
651 volatile cpumask_t other_cpumask
;
654 KKASSERT(&curthread
->td_msgport
== netisr_portfn(0));
655 KKASSERT(!br
->br_isset
);
657 other_cpumask
= mycpu
->gd_other_cpus
& smp_active_mask
;
660 for (i
= 0; i
< ncpus
; ++i
) {
661 struct netmsg_barrier
*msg
;
666 msg
= kmalloc(sizeof(struct netmsg_barrier
),
667 M_LWKTMSG
, M_WAITOK
);
668 netmsg_init(&msg
->base
, NULL
, &netisr_afree_rport
,
669 MSGF_PRIORITY
, netisr_barrier_dispatch
);
670 msg
->br_cpumask
= &other_cpumask
;
671 msg
->br_done
= NETISR_BR_NOTDONE
;
673 KKASSERT(br
->br_msgs
[i
] == NULL
);
674 br
->br_msgs
[i
] = msg
;
677 for (i
= 0; i
< ncpus
; ++i
) {
680 lwkt_sendmsg(netisr_portfn(i
), &br
->br_msgs
[i
]->base
.lmsg
);
683 while (other_cpumask
!= 0) {
684 tsleep_interlock(&other_cpumask
, 0);
685 if (other_cpumask
!= 0)
686 tsleep(&other_cpumask
, PINTERLOCKED
, "nbrset", 0);
693 netisr_barrier_rem(struct netisr_barrier
*br
)
698 KKASSERT(&curthread
->td_msgport
== netisr_portfn(0));
699 KKASSERT(br
->br_isset
);
702 for (i
= 0; i
< ncpus
; ++i
) {
703 struct netmsg_barrier
*msg
= br
->br_msgs
[i
];
706 msg
= br
->br_msgs
[i
];
707 br
->br_msgs
[i
] = NULL
;
712 done
= atomic_swap_int(&msg
->br_done
, 0);
713 if (done
& NETISR_BR_WAITDONE
)
714 wakeup(&msg
->br_done
);
721 netisr_nohashck(struct mbuf
*m
, const struct pktinfo
*pi __unused
)
723 m
->m_flags
&= ~M_HASH
;
727 netisr_hashcheck(int num
, struct mbuf
*m
, const struct pktinfo
*pi
)
731 if (num
< 0 || num
>= NETISR_MAX
)
732 panic("Bad isr %d", num
);
738 if (ni
->ni_handler
== NULL
)
739 panic("Unregistered isr %d", num
);
741 ni
->ni_hashck(m
, pi
);