2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 #include <afsconfig.h>
11 #include "afs/param.h"
14 #include <sys/malloc.h>
15 #include "rx/rx_kcommon.h"
17 #ifdef RXK_LISTENER_ENV
19 osi_NetReceive(osi_socket asocket, struct sockaddr_in *addr,
20 struct iovec *dvec, int nvecs, int *alength)
24 struct iovec iov[RX_MAXIOVECS];
25 struct sockaddr *sa = NULL;
28 int haveGlock = ISAFS_GLOCK();
29 /*AFS_STATCNT(osi_NetReceive); */
31 if (nvecs > RX_MAXIOVECS)
32 osi_Panic("osi_NetReceive: %d: Too many iovecs.\n", nvecs);
34 for (i = 0; i < nvecs; i++)
40 u.uio_resid = *alength;
41 u.uio_segflg = UIO_SYSSPACE;
47 code = soreceive(asocket, &sa, &u, NULL, NULL, NULL);
54 Debugger("afs NetReceive busted");
61 *alength -= u.uio_resid;
63 if (sa->sa_family == AF_INET) {
65 *addr = *(struct sockaddr_in *)sa;
67 printf("Unknown socket family %d in NetReceive\n", sa->sa_family);
73 #define so_is_disconn(so) ((so)->so_state & SS_ISDISCONNECTED)
75 extern int rxk_ListenerPid;
77 osi_StopListener(void)
82 * Have to drop global lock to safely do this.
83 * soclose() is currently protected by Giant,
84 * but pfind and psignal are MPSAFE.
86 int haveGlock = ISAFS_GLOCK();
89 soshutdown(rx_socket, SHUT_RDWR);
90 p = pfind(rxk_ListenerPid);
92 afs_warn("osi_StopListener: rxk_ListenerPid %u\n", rxk_ListenerPid);
96 afs_warn("osi_StopListener: rxk_Listener not found (pid %u)\n",
100 /* Avoid destroying socket until osi_NetReceive has
101 * had a chance to clean up */
105 MUTEX_INIT(&s_mtx, "rx_shutdown_mutex", MUTEX_DEFAULT, 0);
108 while ((tries > 0) && (!so_is_disconn(rx_socket))) {
109 afs_warn("osi_StopListener: waiting (%d) ", tries);
110 msleep(&osi_StopListener, &s_mtx, PSOCK | PCATCH,
111 "rx_shutdown_timedwait", 1 * hz);
116 MUTEX_DESTROY(&s_mtx);
124 osi_NetSend(osi_socket asocket, struct sockaddr_in *addr, struct iovec *dvec,
125 int nvecs, afs_int32 alength, int istack)
129 struct iovec iov[RX_MAXIOVECS];
131 int haveGlock = ISAFS_GLOCK();
133 AFS_STATCNT(osi_NetSend);
134 if (nvecs > RX_MAXIOVECS)
135 osi_Panic("osi_NetSend: %d: Too many iovecs.\n", nvecs);
137 for (i = 0; i < nvecs; i++)
141 u.uio_iovcnt = nvecs;
143 u.uio_resid = alength;
144 u.uio_segflg = UIO_SYSSPACE;
145 u.uio_rw = UIO_WRITE;
148 addr->sin_len = sizeof(struct sockaddr_in);
156 sosend(asocket, (struct sockaddr *)addr, &u, NULL, NULL, 0,
161 Debugger("afs NetSend busted");
171 /* This code *almost* works :( */
172 static struct protosw parent_proto; /* udp proto switch */
173 static void rxk_input(struct mbuf *am, int iphlen);
174 static void rxk_fasttimo(void);
176 /* start intercepting basic calls */
179 struct protosw *tpro, *last;
183 last = inetdomain.dom_protoswNPROTOSW;
184 for (tpro = inetdomain.dom_protosw; tpro < last; tpro++)
185 if (tpro->pr_protocol == IPPROTO_UDP) {
186 #if 0 /* not exported */
187 /* force UDP checksumming on for AFS */
191 memcpy(&parent_proto, tpro, sizeof(parent_proto));
192 tpro->pr_input = rxk_input;
193 tpro->pr_fasttimo = rxk_fasttimo;
195 * don't bother with pr_drain and pr_ctlinput
196 * until we have something to do
201 osi_Panic("inet:no udp");
206 rxk_input(struct mbuf *am, int iphlen)
213 struct udpiphdr *tvu;
217 struct sockaddr_in taddr;
220 int data_len, comp_sum;
225 /* make sure we have base ip and udp headers in first mbuf */
226 if (iphlen > sizeof(struct ip)) {
227 ip_stripoptions(am, NULL);
228 iphlen = sizeof(struct ip);
231 if (am->m_len < sizeof(struct udpiphdr)) {
232 am = m_pullup(am, sizeof(struct udpiphdr));
239 ti = mtod(am, struct ip *);
240 /* skip basic ip hdr */
241 tu = (struct udphdr *)(((char *)ti) + sizeof(struct ip));
243 /* now read the port out */
247 for (tsp = rxk_ports, i = 0; i < MAXRXPORTS; i++) {
248 if (*tsp++ == port) {
249 /* checksum the packet */
251 * Make mbuf data length reflect UDP length.
252 * If not enough data to reflect UDP length, drop.
254 tvu = (struct udpiphdr *)ti;
255 tlen = ntohs((u_short) tvu->ui_ulen);
256 if ((int)ti->ip_len != tlen) {
257 if (tlen > (int)ti->ip_len) {
262 m_adj(am, tlen - (int)ti->ip_len);
264 /* deliver packet to rx */
265 taddr.sin_family = AF_INET; /* compute source address */
266 taddr.sin_port = tu->uh_sport;
267 taddr.sin_addr.s_addr = ti->ip_src.s_addr;
268 taddr.sin_len = sizeof(taddr);
269 tvu = (struct udpiphdr *)ti; /* virtual udp structure, for cksum */
270 /* handle the checksum. Note that this code damages the actual ip
271 * header (replacing it with the virtual one, which is the same size),
272 * so we must ensure we get everything out we need, first */
273 if (tu->uh_sum != 0) {
274 /* if the checksum is there, always check it. It's crazy not
275 * to, unless you can really be sure that your
276 * underlying network (and interfaces and drivers and
277 * DMA hardware, etc!) is error-free. First, fill
278 * in entire virtual ip header. */
279 memset(tvu->ui_i.ih_x1, 0, 9);
280 tvu->ui_len = tvu->ui_ulen;
281 tlen = ntohs((unsigned short)(tvu->ui_ulen));
282 if (in_cksum(am, sizeof(struct ip) + tlen)) {
283 /* checksum, including cksum field, doesn't come out 0, so
284 * this packet is bad */
292 * 28 is IP (20) + UDP (8) header. ulen includes
293 * udp header, and we *don't* tell RX about udp
294 * header either. So, we remove those 8 as well.
296 data_len = ntohs(tu->uh_ulen);
298 if (!(*rxk_GetPacketProc) (&phandle, data_len)) {
299 if (rx_mb_to_packet(am, m_freem, 28, data_len, phandle)) {
300 /* XXX should just increment counter here.. */
301 printf("rx: truncated UDP packet\n");
302 rxi_FreePacket(phandle);
304 (*rxk_PacketArrivalProc) (phandle, &taddr,
305 rxk_portRocks[i], data_len);
314 /* if we get here, try to deliver packet to udp */
315 if (tproc = parent_proto.pr_input)
316 (*tproc) (am, iphlen);
323 * UDP fast timer to raise events for all but Solaris and NCR.
324 * Called about 5 times per second (at unknown priority?). Must go to
325 * splnet or obtain global lock before touching anything significant.
333 /* do rx fasttimo processing here */
334 rxevent_RaiseEvents(&temp);
335 if (tproc = parent_proto.pr_fasttimo)
339 /* rx_NetSend - send asize bytes at adata from asocket to host at addr.
341 * Now, why do we allocate a new buffer when we could theoretically use the one
342 * pointed to by adata? Because PRU_SEND returns after queueing the message,
343 * not after sending it. If the sender changes the data after queueing it,
344 * we'd see the already-queued data change. One attempt to fix this without
345 * adding a copy would be to have this function wait until the datagram is
346 * sent; however this doesn't work well. In particular, if a host is down, and
347 * an ARP fails to that host, this packet will be queued until the ARP request
348 * comes back, which could be hours later. We can't block in this routine that
349 * long, since it prevents RPC timeouts from happening.
351 /* XXX In the brave new world, steal the data bufs out of the rx_packet iovec,
352 * and just queue those. XXX
355 /* set lock on sockbuf sb; can't call sblock since we're at interrupt level
361 AFS_STATCNT(trysblock);
362 if (sb->sb_flags & SB_LOCK) {
363 return -1; /* can't lock socket */
365 sb->sb_flags |= SB_LOCK;
369 /* We only have to do all the mbuf management ourselves if we can be called at
370 interrupt time. in RXK_LISTENER_ENV, we can just call sosend() */
372 osi_NetSend(osi_socket asocket, struct sockaddr_in *addr, struct iovec *dvec,
373 int nvec, afs_int32 asize, int istack)
375 struct mbuf *tm, *um;
378 struct mbuf *top = 0;
379 struct mbuf *m, **mp;
387 static int before = 0;
390 AFS_STATCNT(osi_NetSend);
391 /* Actually, the Ultrix way is as good as any for us, so we don't bother with
392 * special mbufs any more. Used to think we could get away with not copying
393 * the data to the interface, but there's no way to tell the caller not to
394 * reuse the buffers after sending, so we lost out on that trick anyway */
396 if (trysblock(&asocket->so_snd)) {
402 tdata = dvec[i].iov_base;
403 tl = dvec[i].iov_len;
407 MGETHDR(m, M_DONTWAIT, MT_DATA);
409 sbunlock(&asocket->so_snd);
415 m->m_pkthdr.rcvif = NULL;
417 MGET(m, M_DONTWAIT, MT_DATA);
419 /* can't get an mbuf, give up */
421 m_freem(top); /* free mbuf list we're building */
422 sbunlock(&asocket->so_snd);
427 * WARNING: the `4 * MLEN' is somewhat dubious. It is better than
428 * `NBPG', which may have no relation to `CLBYTES'. Also, `CLBYTES'
429 * may be so large that we never use clusters, resulting in far
430 * too many mbufs being used. It is often better to briefly use
431 * a cluster, even if we are only using a portion of it. Since
432 * we are on the xmit side, it shouldn't end up sitting on a queue
433 * for a potentially unbounded time (except perhaps if we are talking
436 if (asize >= 4 * MLEN) { /* try to get cluster mbuf */
437 /* different algorithms for getting cluster mbuf */
438 MCLGET(m, M_DONTWAIT);
439 if ((m->m_flags & M_EXT) == 0)
443 /* now compute usable size */
444 len = MIN(mlen, asize);
445 /* Should I look at MAPPED_MBUFS??? */
448 len = MIN(mlen, asize);
452 top->m_pkthdr.len += len;
453 tpa = mtod(m, caddr_t);
456 memcpy(tpa, tdata, rlen);
466 /* shouldn't come here! */
467 asize = 0; /* so we make progress toward completion */
470 tdata = dvec[i].iov_base;
471 tl = dvec[i].iov_len;
483 /* setup mbuf corresponding to destination address */
484 um = m_get(M_DONTWAIT, MT_SONAME);
487 m_freem(top); /* free mbuf chain */
488 sbunlock(&asocket->so_snd);
492 memcpy(mtod(um, caddr_t), addr, sizeof(*addr));
493 addr->sin_len = um->m_len = sizeof(*addr);
494 /* note that udp_usrreq frees funny mbuf. We hold onto data, but mbuf
495 * around it is gone. */
496 /* haveGlock = ISAFS_GLOCK();
500 /* SOCKET_LOCK(asocket); */
501 /* code = (*asocket->so_proto->pr_usrreq)(asocket, PRU_SEND, tm, um, 0); */
504 Debugger("afs NetSend before");
507 (*asocket->so_proto->pr_usrreqs->pru_send) (asocket, 0, tm,
510 /* SOCKET_UNLOCK(asocket); */
514 sbunlock(&asocket->so_snd);
519 Debugger("afs NetSend busted");