nfs_socket.c

早期freebsd实现

	rep->r_prev->r_next = rep->r_next;
	rep->r_next->r_prev = rep->r_prev;
	splx(s);

	/*
	 * Decrement the outstanding request count.
	 */
	if (rep->r_flags & R_SENT) {
		rep->r_flags &= ~R_SENT;	/* paranoia */
		nmp->nm_sent -= NFS_CWNDSCALE;
	}

	/*
	 * If there was a successful reply and a tprintf msg.
	 * tprintf a response.
	 */
	if (!error && (rep->r_flags & R_TPRINTFMSG))
		nfs_msg(rep->r_procp, nmp->nm_mountp->mnt_stat.f_mntfromname,
		    "is alive again");
	mrep = rep->r_mrep;
	md = rep->r_md;
	dpos = rep->r_dpos;
	if (error) {
		m_freem(rep->r_mreq);
		free((caddr_t)rep, M_NFSREQ);
		return (error);
	}

	/*
	 * break down the rpc header and check if ok
	 */
	nfsm_dissect(tl, u_long *, 3*NFSX_UNSIGNED);
	if (*tl++ == rpc_msgdenied) {
		if (*tl == rpc_mismatch)
			error = EOPNOTSUPP;
		else if ((nmp->nm_flag & NFSMNT_KERB) && *tl++ == rpc_autherr) {
			if (*tl == rpc_rejectedcred && failed_auth == 0) {
				failed_auth++;
				mheadend->m_next = (struct mbuf *)0;
				m_freem(mrep);
				m_freem(rep->r_mreq);
				goto kerbauth;
			} else
				error = EAUTH;
		} else
			error = EACCES;
		m_freem(mrep);
		m_freem(rep->r_mreq);
		free((caddr_t)rep, M_NFSREQ);
		return (error);
	}

	/*
	 * skip over the auth_verf, someday we may want to cache auth_short's
	 * for nfs_reqhead(), but for now just dump it
	 */
	if (*++tl != 0) {
		i = nfsm_rndup(fxdr_unsigned(long, *tl));
		nfsm_adv(i);
	}
	nfsm_dissect(tl, u_long *, NFSX_UNSIGNED);
	/* 0 == ok */
	if (*tl == 0) {
		nfsm_dissect(tl, u_long *, NFSX_UNSIGNED);
		if (*tl != 0) {
			error = fxdr_unsigned(int, *tl);
			m_freem(mrep);
			if ((nmp->nm_flag & NFSMNT_NQNFS) &&
			    error == NQNFS_TRYLATER) {
				error = 0;
				waituntil = time.tv_sec + trylater_delay;
				while (time.tv_sec < waituntil)
					(void) tsleep((caddr_t)&lbolt,
						PSOCK, "nqnfstry", 0);
				trylater_delay *= nfs_backoff[trylater_cnt];
				if (trylater_cnt < 7)
					trylater_cnt++;
				goto tryagain;
			}

			/*
			 * If the File Handle was stale, invalidate the
			 * lookup cache, just in case.
			 */
			if (error == ESTALE)
				cache_purge(vp);
			m_freem(rep->r_mreq);
			free((caddr_t)rep, M_NFSREQ);
			return (error);
		}

		/*
		 * For nqnfs, get any lease in reply
		 */
		if (nmp->nm_flag & NFSMNT_NQNFS) {
			nfsm_dissect(tl, u_long *, NFSX_UNSIGNED);
			if (*tl) {
				np = VTONFS(vp);
				nqlflag = fxdr_unsigned(int, *tl);
				nfsm_dissect(tl, u_long *, 4*NFSX_UNSIGNED);
				cachable = fxdr_unsigned(int, *tl++);
				reqtime += fxdr_unsigned(int, *tl++);
				if (reqtime > time.tv_sec) {
				    fxdr_hyper(tl, &frev);
				    nqnfs_clientlease(nmp, np, nqlflag,
					cachable, reqtime, frev);
				}
			}
		}
		*mrp = mrep;
		*mdp = md;
		*dposp = dpos;
		m_freem(rep->r_mreq);
		FREE((caddr_t)rep, M_NFSREQ);
		return (0);
	}
	m_freem(mrep);
	m_freem(rep->r_mreq);
	free((caddr_t)rep, M_NFSREQ);
	error = EPROTONOSUPPORT;
nfsmout:
	return (error);
}

/*
 * Generate the rpc reply header
 * siz arg. is used to decide if adding a cluster is worthwhile
 */
nfs_rephead(siz, nd, err, cache, frev, mrq, mbp, bposp)
	int siz;
	struct nfsd *nd;
	int err;
	int cache;
	u_quad_t *frev;
	struct mbuf **mrq;
	struct mbuf **mbp;
	caddr_t *bposp;
{
	register u_long *tl;
	register struct mbuf *mreq;
	caddr_t bpos;
	struct mbuf *mb, *mb2;

	MGETHDR(mreq, M_WAIT, MT_DATA);
	mb = mreq;
	/*
	 * If this is a big reply, use a cluster else
	 * try and leave leading space for the lower level headers.
	 */
	siz += RPC_REPLYSIZ;
	if (siz >= MINCLSIZE) {
		MCLGET(mreq, M_WAIT);
	} else
		mreq->m_data += max_hdr;
	tl = mtod(mreq, u_long *);
	mreq->m_len = 6*NFSX_UNSIGNED;
	bpos = ((caddr_t)tl)+mreq->m_len;
	*tl++ = nd->nd_retxid;
	*tl++ = rpc_reply;
	if (err == ERPCMISMATCH || err == NQNFS_AUTHERR) {
		*tl++ = rpc_msgdenied;
		if (err == NQNFS_AUTHERR) {
			*tl++ = rpc_autherr;
			*tl = rpc_rejectedcred;
			mreq->m_len -= NFSX_UNSIGNED;
			bpos -= NFSX_UNSIGNED;
		} else {
			*tl++ = rpc_mismatch;
			*tl++ = txdr_unsigned(2);
			*tl = txdr_unsigned(2);
		}
	} else {
		*tl++ = rpc_msgaccepted;
		*tl++ = 0;
		*tl++ = 0;
		switch (err) {
		case EPROGUNAVAIL:
			*tl = txdr_unsigned(RPC_PROGUNAVAIL);
			break;
		case EPROGMISMATCH:
			*tl = txdr_unsigned(RPC_PROGMISMATCH);
			nfsm_build(tl, u_long *, 2*NFSX_UNSIGNED);
			*tl++ = txdr_unsigned(2);
			*tl = txdr_unsigned(2);	/* someday 3 */
			break;
		case EPROCUNAVAIL:
			*tl = txdr_unsigned(RPC_PROCUNAVAIL);
			break;
		default:
			*tl = 0;
			if (err != VNOVAL) {
				nfsm_build(tl, u_long *, NFSX_UNSIGNED);
				if (err)
					*tl = txdr_unsigned(nfsrv_errmap[err - 1]);
				else
					*tl = 0;
			}
			break;
		};
	}

	/*
	 * For nqnfs, piggyback lease as requested.
	 */
	if (nd->nd_nqlflag != NQL_NOVAL && err == 0) {
		if (nd->nd_nqlflag) {
			nfsm_build(tl, u_long *, 5*NFSX_UNSIGNED);
			*tl++ = txdr_unsigned(nd->nd_nqlflag);
			*tl++ = txdr_unsigned(cache);
			*tl++ = txdr_unsigned(nd->nd_duration);
			txdr_hyper(frev, tl);
		} else {
			if (nd->nd_nqlflag != 0)
				panic("nqreph");
			nfsm_build(tl, u_long *, NFSX_UNSIGNED);
			*tl = 0;
		}
	}
	*mrq = mreq;
	*mbp = mb;
	*bposp = bpos;
	if (err != 0 && err != VNOVAL)
		nfsstats.srvrpc_errs++;
	return (0);
}

/*
 * Nfs timer routine
 * Scan the nfsreq list and retranmit any requests that have timed out
 * To avoid retransmission attempts on STREAM sockets (in the future) make
 * sure to set the r_retry field to 0 (implies nm_retry == 0).
 */
void
nfs_timer(arg)
	void *arg;
{
	register struct nfsreq *rep;
	register struct mbuf *m;
	register struct socket *so;
	register struct nfsmount *nmp;
	register int timeo;
	static long lasttime = 0;
	int s, error;

	s = splnet();
	for (rep = nfsreqh.r_next; rep != &nfsreqh; rep = rep->r_next) {
		nmp = rep->r_nmp;
		if (rep->r_mrep || (rep->r_flags & R_SOFTTERM))
			continue;
		if (nfs_sigintr(nmp, rep, rep->r_procp)) {
			rep->r_flags |= R_SOFTTERM;
			continue;
		}
		if (rep->r_rtt >= 0) {
			rep->r_rtt++;
			if (nmp->nm_flag & NFSMNT_DUMBTIMR)
				timeo = nmp->nm_timeo;
			else
				timeo = NFS_RTO(nmp, proct[rep->r_procnum]);
			if (nmp->nm_timeouts > 0)
				timeo *= nfs_backoff[nmp->nm_timeouts - 1];
			if (rep->r_rtt <= timeo)
				continue;
			if (nmp->nm_timeouts < 8)
				nmp->nm_timeouts++;
		}
		/*
		 * Check for server not responding
		 */
		if ((rep->r_flags & R_TPRINTFMSG) == 0 &&
		     rep->r_rexmit > nmp->nm_deadthresh) {
			nfs_msg(rep->r_procp,
			    nmp->nm_mountp->mnt_stat.f_mntfromname,
			    "not responding");
			rep->r_flags |= R_TPRINTFMSG;
		}
		if (rep->r_rexmit >= rep->r_retry) {	/* too many */
			nfsstats.rpctimeouts++;
			rep->r_flags |= R_SOFTTERM;
			continue;
		}
		if (nmp->nm_sotype != SOCK_DGRAM) {
			if (++rep->r_rexmit > NFS_MAXREXMIT)
				rep->r_rexmit = NFS_MAXREXMIT;
			continue;
		}
		if ((so = nmp->nm_so) == NULL)
			continue;

		/*
		 * If there is enough space and the window allows..
		 *	Resend it
		 * Set r_rtt to -1 in case we fail to send it now.
		 */
		rep->r_rtt = -1;
		if (sbspace(&so->so_snd) >= rep->r_mreq->m_pkthdr.len &&
		   ((nmp->nm_flag & NFSMNT_DUMBTIMR) ||
		    (rep->r_flags & R_SENT) ||
		    nmp->nm_sent < nmp->nm_cwnd) &&
		   (m = m_copym(rep->r_mreq, 0, M_COPYALL, M_DONTWAIT))){
			if ((nmp->nm_flag & NFSMNT_NOCONN) == 0)
			    error = (*so->so_proto->pr_usrreq)(so, PRU_SEND, m,
			    (struct mbuf *)0, (struct mbuf *)0);
			else
			    error = (*so->so_proto->pr_usrreq)(so, PRU_SEND, m,
			    nmp->nm_nam, (struct mbuf *)0);
			if (error) {
				if (NFSIGNORE_SOERROR(nmp->nm_soflags, error))
					so->so_error = 0;
			} else {
				/*
				 * Iff first send, start timing
				 * else turn timing off, backoff timer
				 * and divide congestion window by 2.
				 */
				if (rep->r_flags & R_SENT) {
					rep->r_flags &= ~R_TIMING;
					if (++rep->r_rexmit > NFS_MAXREXMIT)
						rep->r_rexmit = NFS_MAXREXMIT;
					nmp->nm_cwnd >>= 1;
					if (nmp->nm_cwnd < NFS_CWNDSCALE)
						nmp->nm_cwnd = NFS_CWNDSCALE;
					nfsstats.rpcretries++;
				} else {
					rep->r_flags |= R_SENT;
					nmp->nm_sent += NFS_CWNDSCALE;
				}
				rep->r_rtt = 0;
			}
		}
	}

	/*
	 * Call the nqnfs server timer once a second to handle leases.
	 */
	if (lasttime != time.tv_sec) {
		lasttime = time.tv_sec;
		nqnfs_serverd();
	}
	splx(s);
	timeout(nfs_timer, (void *)0, hz / NFS_HZ);
}

/*
 * Test for a termination condition pending on the process.
 * This is used for NFSMNT_INT mounts.
 */
nfs_sigintr(nmp, rep, p)
	struct nfsmount *nmp;
	struct nfsreq *rep;
	register struct proc *p;
{

	if (rep && (rep->r_flags & R_SOFTTERM))
		return (EINTR);
	if (!(nmp->nm_flag & NFSMNT_INT))
		return (0);
	if (p && p->p_siglist &&
	    (((p->p_siglist & ~p->p_sigmask) & ~p->p_sigignore) &
	    NFSINT_SIGMASK))
		return (EINTR);
	return (0);
}

/*
 * Lock a socket against others.
 * Necessary for STREAM sockets to ensure you get an entire rpc request/reply
 * and also to avoid race conditions between the processes with nfs requests
 * in progress when a reconnect is necessary.
 */
nfs_sndlock(flagp, rep)
	register int *flagp;
	struct nfsreq *rep;
{
	struct proc *p;
	int slpflag = 0, slptimeo = 0;

	if (rep) {
		p = rep->r_procp;
		if (rep->r_nmp->nm_flag & NFSMNT_INT)
			slpflag = PCATCH;
	} else
		p = (struct proc *)0;
	while (*flagp & NFSMNT_SNDLOCK) {
		if (nfs_sigintr(rep->r_nmp, rep, p))
			return (EINTR);
		*flagp |= NFSMNT_WANTSND;
		(void) tsleep((caddr_t)flagp, slpflag | (PZERO - 1), "nfsndlck",
			slptimeo);
		if (slpflag == PCATCH) {
			slpflag = 0;
			slptimeo = 2 * hz;
		}
	}
	*flagp |= NFSMNT_SNDLOCK;
	return (0);
}

/*
 * Unlock the stream socket for others.
 */
void
nfs_sndunlock(flagp)
	register int *flagp;
{

	if ((*flagp & NFSMNT_SNDLOCK) == 0)
		panic("nfs sndunlock");
	*flagp &= ~NFSMNT_SNDLOCK;
	if (*flagp & NFSMNT_WANTSND) {
		*flagp &= ~NFSMNT_WANTSND;
		wakeup((caddr_t)flagp);
	}
}

nfs_rcvlock(rep)
	register struct nfsreq *rep;
{
	register int *flagp = &rep->r_nmp->nm_flag;
	int slpflag, slptimeo = 0;

	if (*flagp & NFSMNT_INT)
		slpflag = PCATCH;
	else
		slpflag = 0;
	while (*flagp & NFSMNT_RCVLOCK) {
		if (nfs_sigintr(rep->r_nmp, rep, rep->r_procp))
			return (EINTR);
		*flagp |= NFSMNT_WANTRCV;
		(void) tsleep((caddr_t)flagp, slpflag | (PZERO - 1), "nfsrcvlk",
			slptimeo);
		if (slpflag == PCATCH) {
			slpflag = 0;
			slptimeo = 2 * hz;
		}
	}
	*flagp |= NFSMNT_RCVLOCK;
	return (0);
}

/*
 * Unlock the stream socket for others.
 */
void
nfs_rcvunlock(flagp)
	register int *flagp;
{

	if ((*flagp & NFSMNT_RCVLOCK) == 0)
		panic("nfs rcvunlock");
	*flagp &= ~NFSMNT_RCVLOCK;
	if (*flagp & NFSMNT_WANTRCV) {
		*flagp &= ~NFSMNT_WANTRCV;
		wakeup((caddr_t)flagp);
	}
}

/*
 * Check for badly aligned mbuf data areas and
 * realign data in an mbuf list by copying the data areas up, as required.
 */
void
nfs_realign(m, hsiz)
	register struct mbuf *m;
	int hsiz;
{
	register struct mbuf *m2;
	register int siz, mlen, olen;
	register caddr_t tcp, fcp;
	struct mbuf *mnew;

	while (m) {
	    /*
	     * This never happens for UDP, rarely happens for TCP
	     * but frequently happens for iso transport.
	     */
	    if ((m->m_len & 0x3) || (mtod(m, int) & 0x3)) {
		olen = m->m_len;
		fcp = mtod(m, caddr_t);
		if ((int)fcp & 0x3) {
			m->m_flags &= ~M_PKTHDR;
			if (m->m_flags & M_EXT)
				m->m_data = m->m_ext.ext_buf +
					((m->m_ext.ext_size - olen) & ~0x3);
			else
				m->m_data = m->m_dat;
		}
		m->m_len = 0;
		tcp = mtod(m, caddr_t);
		mnew = m;
		m2 = m->m_next;
	
		/*
		 * If possible, only put the first invariant part
		 * of the RPC header in the first mbuf.
		 */
		mlen = M_TRAILINGSPACE(m);
		if (olen <= hsiz && mlen > hsiz)