2 * Copyright (C) 2013-2014 Universita` di Pisa. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/types.h>
29 #include <sys/module.h>
30 #include <sys/errno.h>
31 #include <sys/param.h> /* defines used in kernel.h */
32 #include <sys/poll.h> /* POLLIN, POLLOUT */
33 #include <sys/kernel.h> /* types used in module initialization */
34 #include <sys/conf.h> /* DEV_MODULE */
35 #include <sys/endian.h>
37 #include <sys/rwlock.h>
39 #include <vm/vm.h> /* vtophys */
40 #include <vm/pmap.h> /* vtophys */
41 #include <vm/vm_param.h>
42 #include <vm/vm_object.h>
43 #include <vm/vm_page.h>
44 #include <vm/vm_pager.h>
48 #include <sys/malloc.h>
49 #include <sys/socket.h> /* sockaddrs */
50 #include <sys/selinfo.h>
52 #include <net/if_var.h>
53 #include <net/if_types.h> /* IFT_ETHER */
54 #include <net/ethernet.h> /* ether_ifdetach */
55 #include <net/if_dl.h> /* LLADDR */
56 #include <machine/bus.h> /* bus_dmamap_* */
57 #include <netinet/in.h> /* in6_cksum_pseudo() */
58 #include <machine/in_cksum.h> /* in_pseudo(), in_cksum_hdr() */
60 #include <net/netmap.h>
61 #include <dev/netmap/netmap_kern.h>
62 #include <dev/netmap/netmap_mem2.h>
65 /* ======================== FREEBSD-SPECIFIC ROUTINES ================== */
68 nm_csum_raw(uint8_t *data, size_t len, rawsum_t cur_sum)
70 /* TODO XXX please use the FreeBSD implementation for this. */
71 uint16_t *words = (uint16_t *)data;
75 for (i = 0; i < nw; i++)
76 cur_sum += be16toh(words[i]);
79 cur_sum += (data[len-1] << 8);
84 /* Fold a raw checksum: 'cur_sum' is in host byte order, while the
85 * return value is in network byte order.
88 nm_csum_fold(rawsum_t cur_sum)
90 /* TODO XXX please use the FreeBSD implementation for this. */
92 cur_sum = (cur_sum & 0xFFFF) + (cur_sum >> 16);
94 return htobe16((~cur_sum) & 0xFFFF);
97 uint16_t nm_csum_ipv4(struct nm_iphdr *iph)
100 return in_cksum_hdr((void *)iph);
102 return nm_csum_fold(nm_csum_raw((uint8_t*)iph, sizeof(struct nm_iphdr), 0));
107 nm_csum_tcpudp_ipv4(struct nm_iphdr *iph, void *data,
108 size_t datalen, uint16_t *check)
111 uint16_t pseudolen = datalen + iph->protocol;
113 /* Compute and insert the pseudo-header cheksum. */
114 *check = in_pseudo(iph->saddr, iph->daddr,
116 /* Compute the checksum on TCP/UDP header + payload
117 * (includes the pseudo-header).
119 *check = nm_csum_fold(nm_csum_raw(data, datalen, 0));
121 static int notsupported = 0;
124 D("inet4 segmentation not supported");
130 nm_csum_tcpudp_ipv6(struct nm_ipv6hdr *ip6h, void *data,
131 size_t datalen, uint16_t *check)
134 *check = in6_cksum_pseudo((void*)ip6h, datalen, ip6h->nexthdr, 0);
135 *check = nm_csum_fold(nm_csum_raw(data, datalen, 0));
137 static int notsupported = 0;
140 D("inet6 segmentation not supported");
147 * Intercept the rx routine in the standard device driver.
148 * Second argument is non-zero to intercept, 0 to restore
151 netmap_catch_rx(struct netmap_adapter *na, int intercept)
153 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
154 struct ifnet *ifp = na->ifp;
157 if (gna->save_if_input) {
158 D("cannot intercept again");
159 return EINVAL; /* already set */
161 gna->save_if_input = ifp->if_input;
162 ifp->if_input = generic_rx_handler;
164 if (!gna->save_if_input){
166 return EINVAL; /* not saved */
168 ifp->if_input = gna->save_if_input;
169 gna->save_if_input = NULL;
177 * Intercept the packet steering routine in the tx path,
178 * so that we can decide which queue is used for an mbuf.
179 * Second argument is non-zero to intercept, 0 to restore.
180 * On freebsd we just intercept if_transmit.
183 netmap_catch_tx(struct netmap_generic_adapter *gna, int enable)
185 struct netmap_adapter *na = &gna->up.up;
186 struct ifnet *ifp = na->ifp;
189 na->if_transmit = ifp->if_transmit;
190 ifp->if_transmit = netmap_transmit;
192 ifp->if_transmit = na->if_transmit;
198 * Transmit routine used by generic_netmap_txsync(). Returns 0 on success
199 * and non-zero on error (which may be packet drops or other errors).
200 * addr and len identify the netmap buffer, m is the (preallocated)
201 * mbuf to use for transmissions.
203 * We should add a reference to the mbuf so the m_freem() at the end
204 * of the transmission does not consume resources.
206 * On FreeBSD, and on multiqueue cards, we can force the queue using
207 * if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
208 * i = m->m_pkthdr.flowid % adapter->num_queues;
210 * i = curcpu % adapter->num_queues;
214 generic_xmit_frame(struct ifnet *ifp, struct mbuf *m,
215 void *addr, u_int len, u_int ring_nr)
220 * The mbuf should be a cluster from our special pool,
221 * so we do not need to do an m_copyback but just copy
222 * (and eventually, just reference the netmap buffer)
225 if (GET_MBUF_REFCNT(m) != 1) {
226 D("invalid refcnt %d for %p",
227 GET_MBUF_REFCNT(m), m);
228 panic("in generic_xmit_frame");
230 // XXX the ext_size check is unnecessary if we link the netmap buf
231 if (m->m_ext.ext_size < len) {
232 RD(5, "size %d < len %d", m->m_ext.ext_size, len);
233 len = m->m_ext.ext_size;
235 if (0) { /* XXX seems to have negligible benefits */
236 m->m_ext.ext_buf = m->m_data = addr;
238 bcopy(addr, m->m_data, len);
240 m->m_len = m->m_pkthdr.len = len;
241 // inc refcount. All ours, we could skip the atomic
242 atomic_fetchadd_int(PNT_MBUF_REFCNT(m), 1);
243 M_HASHTYPE_SET(m, M_HASHTYPE_OPAQUE);
244 m->m_pkthdr.flowid = ring_nr;
245 m->m_pkthdr.rcvif = ifp; /* used for tx notification */
246 ret = NA(ifp)->if_transmit(ifp, m);
251 #if __FreeBSD_version >= 1100005
252 struct netmap_adapter *
253 netmap_getna(if_t ifp)
255 return (NA((struct ifnet *)ifp));
257 #endif /* __FreeBSD_version >= 1100005 */
260 * The following two functions are empty until we have a generic
261 * way to extract the info from the ifp
264 generic_find_num_desc(struct ifnet *ifp, unsigned int *tx, unsigned int *rx)
266 D("called, in tx %d rx %d", *tx, *rx);
272 generic_find_num_queues(struct ifnet *ifp, u_int *txq, u_int *rxq)
274 D("called, in txq %d rxq %d", *txq, *rxq);
275 *txq = netmap_generic_rings;
276 *rxq = netmap_generic_rings;
281 netmap_mitigation_init(struct nm_generic_mit *mit, int idx, struct netmap_adapter *na)
284 mit->mit_pending = 0;
285 mit->mit_ring_idx = idx;
291 netmap_mitigation_start(struct nm_generic_mit *mit)
298 netmap_mitigation_restart(struct nm_generic_mit *mit)
305 netmap_mitigation_active(struct nm_generic_mit *mit)
313 netmap_mitigation_cleanup(struct nm_generic_mit *mit)
319 nm_vi_dummy(struct ifnet *ifp, u_long cmd, caddr_t addr)
325 nm_vi_start(struct ifnet *ifp)
327 panic("nm_vi_start() must not be called");
331 * Index manager of persistent virtual interfaces.
332 * It is used to decide the lowest byte of the MAC address.
333 * We use the same algorithm with management of bridge port index.
335 #define NM_VI_MAX 255
337 uint8_t index[NM_VI_MAX]; /* XXX just for a reasonable number */
343 nm_vi_init_index(void)
346 for (i = 0; i < NM_VI_MAX; i++)
347 nm_vi_indices.index[i] = i;
348 nm_vi_indices.active = 0;
349 mtx_init(&nm_vi_indices.lock, "nm_vi_indices_lock", NULL, MTX_DEF);
352 /* return -1 if no index available */
354 nm_vi_get_index(void)
358 mtx_lock(&nm_vi_indices.lock);
359 ret = nm_vi_indices.active == NM_VI_MAX ? -1 :
360 nm_vi_indices.index[nm_vi_indices.active++];
361 mtx_unlock(&nm_vi_indices.lock);
366 nm_vi_free_index(uint8_t val)
370 mtx_lock(&nm_vi_indices.lock);
371 lim = nm_vi_indices.active;
372 for (i = 0; i < lim; i++) {
373 if (nm_vi_indices.index[i] == val) {
374 /* swap index[lim-1] and j */
375 int tmp = nm_vi_indices.index[lim-1];
376 nm_vi_indices.index[lim-1] = val;
377 nm_vi_indices.index[i] = tmp;
378 nm_vi_indices.active--;
382 if (lim == nm_vi_indices.active)
383 D("funny, index %u didn't found", val);
384 mtx_unlock(&nm_vi_indices.lock);
389 * Implementation of a netmap-capable virtual interface that
390 * registered to the system.
391 * It is based on if_tap.c and ip_fw_log.c in FreeBSD 9.
393 * Note: Linux sets refcount to 0 on allocation of net_device,
394 * then increments it on registration to the system.
395 * FreeBSD sets refcount to 1 on if_alloc(), and does not
396 * increment this refcount on if_attach().
399 nm_vi_persist(const char *name, struct ifnet **ret)
403 uint32_t macaddr_mid;
405 int unit = nm_vi_get_index(); /* just to decide MAC address */
410 * We use the same MAC address generation method with tap
411 * except for the highest octet is 00:be instead of 00:bd
413 macaddr_hi = htons(0x00be); /* XXX tap + 1 */
414 macaddr_mid = (uint32_t) ticks;
415 bcopy(&macaddr_hi, eaddr, sizeof(short));
416 bcopy(&macaddr_mid, &eaddr[2], sizeof(uint32_t));
417 eaddr[5] = (uint8_t)unit;
419 ifp = if_alloc(IFT_ETHER);
421 D("if_alloc failed");
424 if_initname(ifp, name, IF_DUNIT_NONE);
426 ifp->if_flags = IFF_UP | IFF_SIMPLEX | IFF_MULTICAST;
427 ifp->if_init = (void *)nm_vi_dummy;
428 ifp->if_ioctl = nm_vi_dummy;
429 ifp->if_start = nm_vi_start;
430 ifp->if_mtu = ETHERMTU;
431 IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen);
432 ifp->if_capabilities |= IFCAP_LINKSTATE;
433 ifp->if_capenable |= IFCAP_LINKSTATE;
435 ether_ifattach(ifp, eaddr);
439 /* unregister from the system and drop the final refcount */
441 nm_vi_detach(struct ifnet *ifp)
443 nm_vi_free_index(((char *)IF_LLADDR(ifp))[5]);
449 * In order to track whether pages are still mapped, we hook into
450 * the standard cdev_pager and intercept the constructor and
454 struct netmap_vm_handle_t {
456 struct netmap_priv_d *priv;
461 netmap_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
462 vm_ooffset_t foff, struct ucred *cred, u_short *color)
464 struct netmap_vm_handle_t *vmh = handle;
467 D("handle %p size %jd prot %d foff %jd",
468 handle, (intmax_t)size, prot, (intmax_t)foff);
477 netmap_dev_pager_dtor(void *handle)
479 struct netmap_vm_handle_t *vmh = handle;
480 struct cdev *dev = vmh->dev;
481 struct netmap_priv_d *priv = vmh->priv;
484 D("handle %p", handle);
492 netmap_dev_pager_fault(vm_object_t object, vm_ooffset_t offset,
493 int prot, vm_page_t *mres)
495 struct netmap_vm_handle_t *vmh = object->handle;
496 struct netmap_priv_d *priv = vmh->priv;
499 vm_memattr_t memattr;
502 ND("object %p offset %jd prot %d mres %p",
503 object, (intmax_t)offset, prot, mres);
504 memattr = object->memattr;
505 pidx = OFF_TO_IDX(offset);
506 paddr = netmap_mem_ofstophys(priv->np_mref, offset);
508 return VM_PAGER_FAIL;
510 if (((*mres)->flags & PG_FICTITIOUS) != 0) {
512 * If the passed in result page is a fake page, update it with
513 * the new physical address.
516 vm_page_updatefake(page, paddr, memattr);
519 * Replace the passed in reqpage page with our own fake page and
520 * free up the all of the original pages.
522 #ifndef VM_OBJECT_WUNLOCK /* FreeBSD < 10.x */
523 #define VM_OBJECT_WUNLOCK VM_OBJECT_UNLOCK
524 #define VM_OBJECT_WLOCK VM_OBJECT_LOCK
525 #endif /* VM_OBJECT_WUNLOCK */
527 VM_OBJECT_WUNLOCK(object);
528 page = vm_page_getfake(paddr, memattr);
529 VM_OBJECT_WLOCK(object);
532 vm_page_unlock(*mres);
534 vm_page_insert(page, object, pidx);
536 page->valid = VM_PAGE_BITS_ALL;
537 return (VM_PAGER_OK);
541 static struct cdev_pager_ops netmap_cdev_pager_ops = {
542 .cdev_pg_ctor = netmap_dev_pager_ctor,
543 .cdev_pg_dtor = netmap_dev_pager_dtor,
544 .cdev_pg_fault = netmap_dev_pager_fault,
549 netmap_mmap_single(struct cdev *cdev, vm_ooffset_t *foff,
550 vm_size_t objsize, vm_object_t *objp, int prot)
553 struct netmap_vm_handle_t *vmh;
554 struct netmap_priv_d *priv;
558 D("cdev %p foff %jd size %jd objp %p prot %d", cdev,
559 (intmax_t )*foff, (intmax_t )objsize, objp, prot);
561 vmh = malloc(sizeof(struct netmap_vm_handle_t), M_DEVBUF,
568 error = devfs_get_cdevpriv((void**)&priv);
575 error = netmap_get_memory(priv);
579 obj = cdev_pager_allocate(vmh, OBJT_DEVICE,
580 &netmap_cdev_pager_ops, objsize, prot,
583 D("cdev_pager_allocate failed");
602 // XXX can we remove this ?
604 netmap_close(struct cdev *dev, int fflag, int devtype, struct thread *td)
607 D("dev %p fflag 0x%x devtype %d td %p",
608 dev, fflag, devtype, td);
614 netmap_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
616 struct netmap_priv_d *priv;
624 // XXX wait or nowait ?
625 priv = malloc(sizeof(struct netmap_priv_d), M_DEVBUF,
630 error = devfs_set_cdevpriv(priv, netmap_dtor);
634 priv->np_refcount = 1;
639 /******************** kqueue support ****************/
642 * The OS_selwakeup also needs to issue a KNOTE_UNLOCKED.
643 * We use a non-zero argument to distinguish the call from the one
644 * in kevent_scan() which instead also needs to run netmap_poll().
645 * The knote uses a global mutex for the time being. We might
646 * try to reuse the one in the si, but it is not allocated
647 * permanently so it might be a bit tricky.
649 * The *kqfilter function registers one or another f_event
650 * depending on read or write mode.
651 * In the call to f_event() td_fpop is NULL so any child function
652 * calling devfs_get_cdevpriv() would fail - and we need it in
653 * netmap_poll(). As a workaround we store priv into kn->kn_hook
654 * and pass it as first argument to netmap_poll(), which then
655 * uses the failure to tell that we are called from f_event()
656 * and do not need the selrecord().
661 freebsd_selwakeup(struct nm_selinfo *si, int pri)
664 D("on knote %p", &si->si.si_note);
665 selwakeuppri(&si->si, pri);
666 /* use a non-zero hint to tell the notification from the
667 * call done in kqueue_scan() which uses 0
669 KNOTE_UNLOCKED(&si->si.si_note, 0x100 /* notification */);
673 netmap_knrdetach(struct knote *kn)
675 struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook;
676 struct selinfo *si = &priv->np_rxsi->si;
678 D("remove selinfo %p", si);
679 knlist_remove(&si->si_note, kn, 0);
683 netmap_knwdetach(struct knote *kn)
685 struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook;
686 struct selinfo *si = &priv->np_txsi->si;
688 D("remove selinfo %p", si);
689 knlist_remove(&si->si_note, kn, 0);
693 * callback from notifies (generated externally) and our
694 * calls to kevent(). The former we just return 1 (ready)
695 * since we do not know better.
696 * In the latter we call netmap_poll and return 0/1 accordingly.
699 netmap_knrw(struct knote *kn, long hint, int events)
701 struct netmap_priv_d *priv;
705 ND(5, "call from notify");
706 return 1; /* assume we are ready */
709 /* the notification may come from an external thread,
710 * in which case we do not want to run the netmap_poll
711 * This should be filtered above, but check just in case.
713 if (curthread != priv->np_td) { /* should not happen */
714 RD(5, "curthread changed %p %p", curthread, priv->np_td);
717 revents = netmap_poll((void *)priv, events, curthread);
718 return (events & revents) ? 1 : 0;
723 netmap_knread(struct knote *kn, long hint)
725 return netmap_knrw(kn, hint, POLLIN);
729 netmap_knwrite(struct knote *kn, long hint)
731 return netmap_knrw(kn, hint, POLLOUT);
734 static struct filterops netmap_rfiltops = {
736 .f_detach = netmap_knrdetach,
737 .f_event = netmap_knread,
740 static struct filterops netmap_wfiltops = {
742 .f_detach = netmap_knwdetach,
743 .f_event = netmap_knwrite,
748 * This is called when a thread invokes kevent() to record
749 * a change in the configuration of the kqueue().
750 * The 'priv' should be the same as in the netmap device.
753 netmap_kqfilter(struct cdev *dev, struct knote *kn)
755 struct netmap_priv_d *priv;
757 struct netmap_adapter *na;
758 struct nm_selinfo *si;
759 int ev = kn->kn_filter;
761 if (ev != EVFILT_READ && ev != EVFILT_WRITE) {
762 D("bad filter request %d", ev);
765 error = devfs_get_cdevpriv((void**)&priv);
767 D("device not yet setup");
772 D("no netmap adapter for this file descriptor");
775 /* the si is indicated in the priv */
776 si = (ev == EVFILT_WRITE) ? priv->np_txsi : priv->np_rxsi;
778 kn->kn_fop = (ev == EVFILT_WRITE) ?
779 &netmap_wfiltops : &netmap_rfiltops;
781 knlist_add(&si->si.si_note, kn, 1);
783 ND("register %p %s td %p priv %p kn %p np_nifp %p kn_fp/fpop %s",
784 na, na->ifp->if_xname, curthread, priv, kn,
786 kn->kn_fp == curthread->td_fpop ? "match" : "MISMATCH");
790 struct cdevsw netmap_cdevsw = {
791 .d_version = D_VERSION,
793 .d_open = netmap_open,
794 .d_mmap_single = netmap_mmap_single,
795 .d_ioctl = netmap_ioctl,
796 .d_poll = netmap_poll,
797 .d_kqfilter = netmap_kqfilter,
798 .d_close = netmap_close,
800 /*--- end of kqueue support ----*/
803 * Kernel entry point.
805 * Initialize/finalize the module and return.
807 * Return 0 on success, errno on failure.
810 netmap_loader(__unused struct module *module, int event, __unused void *arg)
816 error = netmap_init();
832 DEV_MODULE(netmap, netmap_loader, NULL);