Index | Thread | Search

From:
Vitaliy Makkoveev <otto@bsdbox.dev>
Subject:
Re: run TCP syn cache timer in parallel
To:
Alexander Bluhm <alexander.bluhm@gmx.net>
Cc:
OpenBSD tech <tech@openbsd.org>
Date:
Thu, 2 Jan 2025 08:29:27 +0300

Download raw body.

Thread
> On 1 Jan 2025, at 21:39, Alexander Bluhm <alexander.bluhm@gmx.net> wrote:
> 
> On Wed, Jan 01, 2025 at 06:02:54PM +0300, Vitaliy Makkoveev wrote:
>>> On 1 Jan 2025, at 16:01, Alexander Bluhm <bluhm@openbsd.org> wrote:
>>> 
>>> Hi,
>>> 
>>> With mvs@'s socket ref counting, I think it is possible to run TCP
>>> SYN cache in parallel.
>>> 
>>> Idea is to reset inp_socket pointer when socket is about to be
>>> freed.  By protecting this with a mutex, we reliable retrieve the
>>> socket from an inpcb reference.  By using the socket refcount we
>>> switch from mutex to socket lock.
>>> 
>>> The syn cache timer is an easy place to use this.  Goal it to use
>>> in_pcbsolock() also for TCP timer and input.  Note that switching
>>> sc_tp to sc_inplisten and refcounting it, is a diff I sent to tech@
>>> earlier.
>>> 
>>> ok?
>>> 
>> 
>> Hi Alexander,
>> 
>> This diff lokks good to me. However, I don???t like the sorele()
>> after so_lock rwlock(9) release. I made this with previous
>> sorele() and this worked perfectly because exclusive solock()
>> should be taken before continue teardown, but now I prefer to keep
>> the reference until we leave this socket alone. This doesn???t bring
>> any extra impact.
>> 
>> So I like to rename in_pcbsolock() to in_pcb_solock_ref() and
>> in_pcbsounlock() to in_pcb_sounlock_rele() and rework them as
>> something like below.
>> 
>> in_pcb_solock_ref()
>> {
>> 	mtx_enter(&inp->inp_sofree_mtx);
>> 	so = soref(inp->inp_socket);
>> 	mtx_leave(&inp->inp_sofree_mtx);
>> 
>> 	if (so)
>> 		rw_enter_write(&so->so_lock);
>> 	return so;
>> }
>> 
>> in_pcb_sounlock_rele()
>> {
>> 	rw_exit_write(&so->so_lock);
>> 	sorele(so, 1);
>> }
> 
> I have an experimental diff with in_pcbsolock() that runs tcp_input()
> in parallel.  It seems that keeping locks and refcounts gets more
> complicated than just relying on so_lock alone to avoid destruction
> in a different thread.
> 
> So I prefer to keep in_pcbsolock() as it is.  We can always change
> it later.
> 

Go ahead, we always could correct usage with the new vision.

> First step would be to get the sc_inplisten refcount part commited.
> 
> bluhm
> 

I’m working on sockets splice unlocking, and I have the cases, then
socket lock was taken and released before reference release.
sorele() does nothing that requires solock() to be taken, so I think
it’s normal to allow sorele() be called unlocked. We could rename
keep_lock arg to something like no_unlock for consistency. As you
can see there is nothing complicated:

                sblock(&so->so_rcv, SBL_WAIT | SBL_NOINTR);
                if (isspliced(so)) {
                        struct socket *sosp = so->so_sp->ssp_socket;
                        int freeing = SOSP_FREEING_READ;

                        if (so == so->so_sp->ssp_socket)
                                freeing |= SOSP_FREEING_WRITE;
                        soref(sosp);
                        sounsplice(so, so->so_sp->ssp_socket, freeing);
                        sorele(sosp, 1);
                }
                sbunlock(&so->so_rcv);

sounsplice(struct socket *so, struct socket *sosp, int freeing)
{               
        sbassertlocked(&so->so_rcv);
	
	/* Do unsplicing */

        /* Do not wakeup a socket that is about to be freed. */
        if ((freeing & SOSP_FREEING_READ) == 0) {
                solock_shared(so);
                if (soreadable(so))
                        sorwakeup(so);
                sounlock_shared(so);
        }
        if ((freeing & SOSP_FREEING_WRITE) == 0) {
                solock_shared(sosp);
                if (sowriteable(sosp))
                        sowwakeup(sosp);
                sounlock_shared(sosp);
        }
}


>>> Index: netinet/in_pcb.c
>>> ===================================================================
>>> RCS file: /data/mirror/openbsd/cvs/src/sys/netinet/in_pcb.c,v
>>> diff -u -p -r1.307 in_pcb.c
>>> --- netinet/in_pcb.c	24 Dec 2024 16:27:07 -0000	1.307
>>> +++ netinet/in_pcb.c	1 Jan 2025 12:55:13 -0000
>>> @@ -584,6 +584,9 @@ in_pcbdetach(struct inpcb *inp)
>>> 	struct inpcbtable *table = inp->inp_table;
>>> 
>>> 	so->so_pcb = NULL;
>>> +	mtx_enter(&inp->inp_sofree_mtx);
>>> +	inp->inp_socket = NULL;
>>> +	mtx_leave(&inp->inp_sofree_mtx);
>>> 	/*
>>> 	 * As long as the NET_LOCK() is the default lock for Internet
>>> 	 * sockets, do not release it to not introduce new sleeping
>>> @@ -616,6 +619,32 @@ in_pcbdetach(struct inpcb *inp)
>>> 	mtx_leave(&table->inpt_mtx);
>>> 
>>> 	in_pcbunref(inp);
>>> +}
>>> +
>>> +struct socket *
>>> +in_pcbsolock(struct inpcb *inp)
>>> +{
>>> +	struct socket *so;
>>> +
>>> +	NET_ASSERT_LOCKED();
>>> +
>>> +	mtx_enter(&inp->inp_sofree_mtx);
>>> +	so = soref(inp->inp_socket);
>>> +	mtx_leave(&inp->inp_sofree_mtx);
>>> +	if (so == NULL)
>>> +		return NULL;
>>> +
>>> +	rw_enter_write(&so->so_lock);
>>> +	sorele(so, 1);
>>> +
>>> +	return so;
>>> +}
>>> +
>>> +void
>>> +in_pcbsounlock(struct inpcb *inp, struct socket *so)
>>> +{
>>> +	KASSERT(inp->inp_socket == so);
>>> +	rw_exit_write(&so->so_lock);
>>> }
>>> 
>>> struct inpcb *
>>> Index: netinet/in_pcb.h
>>> ===================================================================
>>> RCS file: /data/mirror/openbsd/cvs/src/sys/netinet/in_pcb.h,v
>>> diff -u -p -r1.161 in_pcb.h
>>> --- netinet/in_pcb.h	24 Dec 2024 16:27:07 -0000	1.161
>>> +++ netinet/in_pcb.h	1 Jan 2025 10:53:17 -0000
>>> @@ -81,6 +81,7 @@
>>> *	t	inpt_mtx		pcb table mutex
>>> *	L	pf_inp_mtx		link pf to inp mutex
>>> *	s	so_lock			socket rwlock
>>> + *	f	inp_sofree_mtx		socket detach and lock
>>> */
>>> 
>>> /*
>>> @@ -136,7 +137,8 @@ struct inpcb {
>>> #define	inp_laddr6	inp_laddru.iau_addr6
>>> 	u_int16_t inp_fport;		/* [t] foreign port */
>>> 	u_int16_t inp_lport;		/* [t] local port */
>>> -	struct	  socket *inp_socket;	/* [I] back pointer to socket */
>>> +	struct	  socket *inp_socket;	/* [f] back pointer to socket */
>>> +	struct	  mutex inp_sofree_mtx;	/* protect socket free */
>>> 	caddr_t	  inp_ppcb;		/* pointer to per-protocol pcb */
>>> 	struct    route inp_route;	/* [s] cached route */
>>> 	struct    refcnt inp_refcnt;	/* refcount PCB, delay memory free */
>>> @@ -309,6 +311,9 @@ int	 in_pcbaddrisavail(const struct inpc
>>> 	    struct proc *);
>>> int	 in_pcbconnect(struct inpcb *, struct mbuf *);
>>> void	 in_pcbdetach(struct inpcb *);
>>> +struct socket *
>>> +	 in_pcbsolock(struct inpcb *);
>>> +void	 in_pcbsounlock(struct inpcb *, struct socket *);
>>> struct inpcb *
>>> 	 in_pcbref(struct inpcb *);
>>> void	 in_pcbunref(struct inpcb *);
>>> Index: netinet/tcp_input.c
>>> ===================================================================
>>> RCS file: /data/mirror/openbsd/cvs/src/sys/netinet/tcp_input.c,v
>>> diff -u -p -r1.416 tcp_input.c
>>> --- netinet/tcp_input.c	31 Dec 2024 12:19:46 -0000	1.416
>>> +++ netinet/tcp_input.c	1 Jan 2025 10:49:40 -0000
>>> @@ -3173,7 +3173,8 @@ syn_cache_rm(struct syn_cache *sc)
>>> 	KASSERT(!ISSET(sc->sc_dynflags, SCF_DEAD));
>>> 	SET(sc->sc_dynflags, SCF_DEAD);
>>> 	TAILQ_REMOVE(&sc->sc_buckethead->sch_bucket, sc, sc_bucketq);
>>> -	sc->sc_tp = NULL;
>>> +	in_pcbunref(sc->sc_inplisten);
>>> +	sc->sc_inplisten = NULL;
>>> 	LIST_REMOVE(sc, sc_tpq);
>>> 	refcnt_rele(&sc->sc_refcnt);
>>> 	sc->sc_buckethead->sch_length--;
>>> @@ -3369,6 +3370,8 @@ void
>>> syn_cache_timer(void *arg)
>>> {
>>> 	struct syn_cache *sc = arg;
>>> +	struct inpcb *inp;
>>> +	struct socket *so;
>>> 	uint64_t now;
>>> 	int lastref;
>>> 
>>> @@ -3397,14 +3400,22 @@ syn_cache_timer(void *arg)
>>> 	    TCPTV_REXMTMAX);
>>> 	if (timeout_add_msec(&sc->sc_timer, sc->sc_rxtcur))
>>> 		refcnt_take(&sc->sc_refcnt);
>>> +	inp = in_pcbref(sc->sc_inplisten);
>>> +	if (inp == NULL)
>>> +		goto freeit;
>>> 	mtx_leave(&syn_cache_mtx);
>>> 
>>> -	NET_LOCK();
>>> -	now = tcp_now();
>>> -	(void) syn_cache_respond(sc, NULL, now);
>>> -	tcpstat_inc(tcps_sc_retransmitted);
>>> -	NET_UNLOCK();
>>> +	NET_LOCK_SHARED();
>>> +	so = in_pcbsolock(inp);
>>> +	if (so != NULL) {
>>> +		now = tcp_now();
>>> +		(void) syn_cache_respond(sc, NULL, now);
>>> +		tcpstat_inc(tcps_sc_retransmitted);
>>> +		in_pcbsounlock(inp, so);
>>> +	}
>>> +	NET_UNLOCK_SHARED();
>>> 
>>> +	in_pcbunref(inp);
>>> 	syn_cache_put(sc);
>>> 	return;
>>> 
>>> @@ -3434,10 +3445,7 @@ syn_cache_cleanup(struct tcpcb *tp)
>>> 
>>> 	mtx_enter(&syn_cache_mtx);
>>> 	LIST_FOREACH_SAFE(sc, &tp->t_sc, sc_tpq, nsc) {
>>> -#ifdef DIAGNOSTIC
>>> -		if (sc->sc_tp != tp)
>>> -			panic("invalid sc_tp in syn_cache_cleanup");
>>> -#endif
>>> +		KASSERT(sc->sc_inplisten == tp->t_inpcb);
>>> 		syn_cache_rm(sc);
>>> 		syn_cache_put(sc);
>>> 	}
>>> @@ -3949,7 +3957,7 @@ syn_cache_add(struct sockaddr *src, stru
>>> 	if (tb.t_flags & TF_SIGNATURE)
>>> 		SET(sc->sc_fixflags, SCF_SIGNATURE);
>>> #endif
>>> -	sc->sc_tp = tp;
>>> +	sc->sc_inplisten = in_pcbref(tp->t_inpcb);
>>> 	if (syn_cache_respond(sc, m, now) == 0) {
>>> 		mtx_enter(&syn_cache_mtx);
>>> 		/*
>>> @@ -3962,6 +3970,7 @@ syn_cache_add(struct sockaddr *src, stru
>>> 		tcpstat_inc(tcps_sndacks);
>>> 		tcpstat_inc(tcps_sndtotal);
>>> 	} else {
>>> +		in_pcbunref(sc->sc_inplisten);
>>> 		syn_cache_put(sc);
>>> 		tcpstat_inc(tcps_sc_dropped);
>>> 	}
>>> @@ -4158,7 +4167,7 @@ syn_cache_respond(struct syn_cache *sc,
>>> 
>>> 	/* use IPsec policy and ttl from listening socket, on SYN ACK */
>>> 	mtx_enter(&syn_cache_mtx);
>>> -	inp = sc->sc_tp ? sc->sc_tp->t_inpcb : NULL;
>>> +	inp = in_pcbref(sc->sc_inplisten);
>>> 	mtx_leave(&syn_cache_mtx);
>>> 
>>> 	/*
>>> @@ -4189,5 +4198,6 @@ syn_cache_respond(struct syn_cache *sc,
>>> 		break;
>>> #endif
>>> 	}
>>> +	in_pcbunref(inp);
>>> 	return (error);
>>> }
>>> Index: netinet/tcp_var.h
>>> ===================================================================
>>> RCS file: /data/mirror/openbsd/cvs/src/sys/netinet/tcp_var.h,v
>>> diff -u -p -r1.181 tcp_var.h
>>> --- netinet/tcp_var.h	28 Dec 2024 22:17:09 -0000	1.181
>>> +++ netinet/tcp_var.h	1 Jan 2025 10:26:36 -0000
>>> @@ -278,7 +278,7 @@ struct syn_cache {
>>> 	u_int     sc_request_r_scale	: 4,	/* [I] */
>>> 		  sc_requested_s_scale	: 4;	/* [I] */
>>> 
>>> -	struct tcpcb *sc_tp;		/* [S] tcb for listening socket */
>>> +	struct inpcb *sc_inplisten;	/* [S] inpcb for listening socket */
>>> 	LIST_ENTRY(syn_cache) sc_tpq;	/* [S] list of entries by same tp */
>>> };
>>> 
>>> Index: sys/socketvar.h
>>> ===================================================================
>>> RCS file: /data/mirror/openbsd/cvs/src/sys/sys/socketvar.h,v
>>> diff -u -p -r1.135 socketvar.h
>>> --- sys/socketvar.h	30 Dec 2024 12:12:35 -0000	1.135
>>> +++ sys/socketvar.h	1 Jan 2025 10:56:10 -0000
>>> @@ -209,10 +209,13 @@ struct socket {
>>> void	soassertlocked(struct socket *);
>>> void	soassertlocked_readonly(struct socket *);
>>> 
>>> -static inline void
>>> +static inline struct socket *
>>> soref(struct socket *so)
>>> {
>>> +	if (so == NULL)
>>> +		return NULL;
>>> 	refcnt_take(&so->so_refcnt);
>>> +	return so;
>>> }
>>> 
>>> /*
>>> 
>>