netmap_freebsd.c revision 285445
1/* 2 * Copyright (C) 2013-2014 Universita` di Pisa. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 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. 12 * 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 23 * SUCH DAMAGE. 24 */ 25 26/* $FreeBSD: head/sys/dev/netmap/netmap_freebsd.c 285445 2015-07-13 10:23:52Z luigi $ */ 27#include "opt_inet.h" 28#include "opt_inet6.h" 29 30#include <sys/types.h> 31#include <sys/module.h> 32#include <sys/errno.h> 33#include <sys/param.h> /* defines used in kernel.h */ 34#include <sys/poll.h> /* POLLIN, POLLOUT */ 35#include <sys/kernel.h> /* types used in module initialization */ 36#include <sys/conf.h> /* DEV_MODULE */ 37#include <sys/endian.h> 38 39#include <sys/rwlock.h> 40 41#include <vm/vm.h> /* vtophys */ 42#include <vm/pmap.h> /* vtophys */ 43#include <vm/vm_param.h> 44#include <vm/vm_object.h> 45#include <vm/vm_page.h> 46#include <vm/vm_pager.h> 47#include <vm/uma.h> 48 49 50#include <sys/malloc.h> 51#include <sys/socket.h> /* sockaddrs */ 52#include <sys/selinfo.h> 53#include <net/if.h> 54#include <net/if_var.h> 55#include <net/if_types.h> /* IFT_ETHER */ 56#include <net/ethernet.h> /* ether_ifdetach */ 57#include <net/if_dl.h> /* LLADDR */ 58#include <machine/bus.h> /* bus_dmamap_* */ 59#include <netinet/in.h> /* in6_cksum_pseudo() */ 60#include <machine/in_cksum.h> /* in_pseudo(), in_cksum_hdr() */ 61 62#include <net/netmap.h> 63#include <dev/netmap/netmap_kern.h> 64#include <dev/netmap/netmap_mem2.h> 65 66 67/* ======================== FREEBSD-SPECIFIC ROUTINES ================== */ 68 69rawsum_t 70nm_csum_raw(uint8_t *data, size_t len, rawsum_t cur_sum) 71{ 72 /* TODO XXX please use the FreeBSD implementation for this. */ 73 uint16_t *words = (uint16_t *)data; 74 int nw = len / 2; 75 int i; 76 77 for (i = 0; i < nw; i++) 78 cur_sum += be16toh(words[i]); 79 80 if (len & 1) 81 cur_sum += (data[len-1] << 8); 82 83 return cur_sum; 84} 85 86/* Fold a raw checksum: 'cur_sum' is in host byte order, while the 87 * return value is in network byte order. 88 */ 89uint16_t 90nm_csum_fold(rawsum_t cur_sum) 91{ 92 /* TODO XXX please use the FreeBSD implementation for this. */ 93 while (cur_sum >> 16) 94 cur_sum = (cur_sum & 0xFFFF) + (cur_sum >> 16); 95 96 return htobe16((~cur_sum) & 0xFFFF); 97} 98 99uint16_t nm_csum_ipv4(struct nm_iphdr *iph) 100{ 101#if 0 102 return in_cksum_hdr((void *)iph); 103#else 104 return nm_csum_fold(nm_csum_raw((uint8_t*)iph, sizeof(struct nm_iphdr), 0)); 105#endif 106} 107 108void 109nm_csum_tcpudp_ipv4(struct nm_iphdr *iph, void *data, 110 size_t datalen, uint16_t *check) 111{ 112#ifdef INET 113 uint16_t pseudolen = datalen + iph->protocol; 114 115 /* Compute and insert the pseudo-header cheksum. */ 116 *check = in_pseudo(iph->saddr, iph->daddr, 117 htobe16(pseudolen)); 118 /* Compute the checksum on TCP/UDP header + payload 119 * (includes the pseudo-header). 120 */ 121 *check = nm_csum_fold(nm_csum_raw(data, datalen, 0)); 122#else 123 static int notsupported = 0; 124 if (!notsupported) { 125 notsupported = 1; 126 D("inet4 segmentation not supported"); 127 } 128#endif 129} 130 131void 132nm_csum_tcpudp_ipv6(struct nm_ipv6hdr *ip6h, void *data, 133 size_t datalen, uint16_t *check) 134{ 135#ifdef INET6 136 *check = in6_cksum_pseudo((void*)ip6h, datalen, ip6h->nexthdr, 0); 137 *check = nm_csum_fold(nm_csum_raw(data, datalen, 0)); 138#else 139 static int notsupported = 0; 140 if (!notsupported) { 141 notsupported = 1; 142 D("inet6 segmentation not supported"); 143 } 144#endif 145} 146 147 148/* 149 * Intercept the rx routine in the standard device driver. 150 * Second argument is non-zero to intercept, 0 to restore 151 */ 152int 153netmap_catch_rx(struct netmap_generic_adapter *gna, int intercept) 154{ 155 struct netmap_adapter *na = &gna->up.up; 156 struct ifnet *ifp = na->ifp; 157 158 if (intercept) { 159 if (gna->save_if_input) { 160 D("cannot intercept again"); 161 return EINVAL; /* already set */ 162 } 163 gna->save_if_input = ifp->if_input; 164 ifp->if_input = generic_rx_handler; 165 } else { 166 if (!gna->save_if_input){ 167 D("cannot restore"); 168 return EINVAL; /* not saved */ 169 } 170 ifp->if_input = gna->save_if_input; 171 gna->save_if_input = NULL; 172 } 173 174 return 0; 175} 176 177 178/* 179 * Intercept the packet steering routine in the tx path, 180 * so that we can decide which queue is used for an mbuf. 181 * Second argument is non-zero to intercept, 0 to restore. 182 * On freebsd we just intercept if_transmit. 183 */ 184void 185netmap_catch_tx(struct netmap_generic_adapter *gna, int enable) 186{ 187 struct netmap_adapter *na = &gna->up.up; 188 struct ifnet *ifp = netmap_generic_getifp(gna); 189 190 if (enable) { 191 na->if_transmit = ifp->if_transmit; 192 ifp->if_transmit = netmap_transmit; 193 } else { 194 ifp->if_transmit = na->if_transmit; 195 } 196} 197 198 199/* 200 * Transmit routine used by generic_netmap_txsync(). Returns 0 on success 201 * and non-zero on error (which may be packet drops or other errors). 202 * addr and len identify the netmap buffer, m is the (preallocated) 203 * mbuf to use for transmissions. 204 * 205 * We should add a reference to the mbuf so the m_freem() at the end 206 * of the transmission does not consume resources. 207 * 208 * On FreeBSD, and on multiqueue cards, we can force the queue using 209 * if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) 210 * i = m->m_pkthdr.flowid % adapter->num_queues; 211 * else 212 * i = curcpu % adapter->num_queues; 213 * 214 */ 215int 216generic_xmit_frame(struct ifnet *ifp, struct mbuf *m, 217 void *addr, u_int len, u_int ring_nr) 218{ 219 int ret; 220 221 /* 222 * The mbuf should be a cluster from our special pool, 223 * so we do not need to do an m_copyback but just copy 224 * (and eventually, just reference the netmap buffer) 225 */ 226 227 if (GET_MBUF_REFCNT(m) != 1) { 228 D("invalid refcnt %d for %p", 229 GET_MBUF_REFCNT(m), m); 230 panic("in generic_xmit_frame"); 231 } 232 // XXX the ext_size check is unnecessary if we link the netmap buf 233 if (m->m_ext.ext_size < len) { 234 RD(5, "size %d < len %d", m->m_ext.ext_size, len); 235 len = m->m_ext.ext_size; 236 } 237 if (0) { /* XXX seems to have negligible benefits */ 238 m->m_ext.ext_buf = m->m_data = addr; 239 } else { 240 bcopy(addr, m->m_data, len); 241 } 242 m->m_len = m->m_pkthdr.len = len; 243 // inc refcount. All ours, we could skip the atomic 244 atomic_fetchadd_int(PNT_MBUF_REFCNT(m), 1); 245 M_HASHTYPE_SET(m, M_HASHTYPE_OPAQUE); 246 m->m_pkthdr.flowid = ring_nr; 247 m->m_pkthdr.rcvif = ifp; /* used for tx notification */ 248 ret = NA(ifp)->if_transmit(ifp, m); 249 return ret; 250} 251 252 253#if __FreeBSD_version >= 1100005 254struct netmap_adapter * 255netmap_getna(if_t ifp) 256{ 257 return (NA((struct ifnet *)ifp)); 258} 259#endif /* __FreeBSD_version >= 1100005 */ 260 261/* 262 * The following two functions are empty until we have a generic 263 * way to extract the info from the ifp 264 */ 265int 266generic_find_num_desc(struct ifnet *ifp, unsigned int *tx, unsigned int *rx) 267{ 268 D("called, in tx %d rx %d", *tx, *rx); 269 return 0; 270} 271 272 273void 274generic_find_num_queues(struct ifnet *ifp, u_int *txq, u_int *rxq) 275{ 276 D("called, in txq %d rxq %d", *txq, *rxq); 277 *txq = netmap_generic_rings; 278 *rxq = netmap_generic_rings; 279} 280 281 282void 283netmap_mitigation_init(struct nm_generic_mit *mit, int idx, struct netmap_adapter *na) 284{ 285 ND("called"); 286 mit->mit_pending = 0; 287 mit->mit_ring_idx = idx; 288 mit->mit_na = na; 289} 290 291 292void 293netmap_mitigation_start(struct nm_generic_mit *mit) 294{ 295 ND("called"); 296} 297 298 299void 300netmap_mitigation_restart(struct nm_generic_mit *mit) 301{ 302 ND("called"); 303} 304 305 306int 307netmap_mitigation_active(struct nm_generic_mit *mit) 308{ 309 ND("called"); 310 return 0; 311} 312 313 314void 315netmap_mitigation_cleanup(struct nm_generic_mit *mit) 316{ 317 ND("called"); 318} 319 320static int 321nm_vi_dummy(struct ifnet *ifp, u_long cmd, caddr_t addr) 322{ 323 return EINVAL; 324} 325 326static void 327nm_vi_start(struct ifnet *ifp) 328{ 329 panic("nm_vi_start() must not be called"); 330} 331 332/* 333 * Index manager of persistent virtual interfaces. 334 * It is used to decide the lowest byte of the MAC address. 335 * We use the same algorithm with management of bridge port index. 336 */ 337#define NM_VI_MAX 255 338static struct { 339 uint8_t index[NM_VI_MAX]; /* XXX just for a reasonable number */ 340 uint8_t active; 341 struct mtx lock; 342} nm_vi_indices; 343 344void 345nm_vi_init_index(void) 346{ 347 int i; 348 for (i = 0; i < NM_VI_MAX; i++) 349 nm_vi_indices.index[i] = i; 350 nm_vi_indices.active = 0; 351 mtx_init(&nm_vi_indices.lock, "nm_vi_indices_lock", NULL, MTX_DEF); 352} 353 354/* return -1 if no index available */ 355static int 356nm_vi_get_index(void) 357{ 358 int ret; 359 360 mtx_lock(&nm_vi_indices.lock); 361 ret = nm_vi_indices.active == NM_VI_MAX ? -1 : 362 nm_vi_indices.index[nm_vi_indices.active++]; 363 mtx_unlock(&nm_vi_indices.lock); 364 return ret; 365} 366 367static void 368nm_vi_free_index(uint8_t val) 369{ 370 int i, lim; 371 372 mtx_lock(&nm_vi_indices.lock); 373 lim = nm_vi_indices.active; 374 for (i = 0; i < lim; i++) { 375 if (nm_vi_indices.index[i] == val) { 376 /* swap index[lim-1] and j */ 377 int tmp = nm_vi_indices.index[lim-1]; 378 nm_vi_indices.index[lim-1] = val; 379 nm_vi_indices.index[i] = tmp; 380 nm_vi_indices.active--; 381 break; 382 } 383 } 384 if (lim == nm_vi_indices.active) 385 D("funny, index %u didn't found", val); 386 mtx_unlock(&nm_vi_indices.lock); 387} 388#undef NM_VI_MAX 389 390/* 391 * Implementation of a netmap-capable virtual interface that 392 * registered to the system. 393 * It is based on if_tap.c and ip_fw_log.c in FreeBSD 9. 394 * 395 * Note: Linux sets refcount to 0 on allocation of net_device, 396 * then increments it on registration to the system. 397 * FreeBSD sets refcount to 1 on if_alloc(), and does not 398 * increment this refcount on if_attach(). 399 */ 400int 401nm_vi_persist(const char *name, struct ifnet **ret) 402{ 403 struct ifnet *ifp; 404 u_short macaddr_hi; 405 uint32_t macaddr_mid; 406 u_char eaddr[6]; 407 int unit = nm_vi_get_index(); /* just to decide MAC address */ 408 409 if (unit < 0) 410 return EBUSY; 411 /* 412 * We use the same MAC address generation method with tap 413 * except for the highest octet is 00:be instead of 00:bd 414 */ 415 macaddr_hi = htons(0x00be); /* XXX tap + 1 */ 416 macaddr_mid = (uint32_t) ticks; 417 bcopy(&macaddr_hi, eaddr, sizeof(short)); 418 bcopy(&macaddr_mid, &eaddr[2], sizeof(uint32_t)); 419 eaddr[5] = (uint8_t)unit; 420 421 ifp = if_alloc(IFT_ETHER); 422 if (ifp == NULL) { 423 D("if_alloc failed"); 424 return ENOMEM; 425 } 426 if_initname(ifp, name, IF_DUNIT_NONE); 427 ifp->if_mtu = 65536; 428 ifp->if_flags = IFF_UP | IFF_SIMPLEX | IFF_MULTICAST; 429 ifp->if_init = (void *)nm_vi_dummy; 430 ifp->if_ioctl = nm_vi_dummy; 431 ifp->if_start = nm_vi_start; 432 ifp->if_mtu = ETHERMTU; 433 IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); 434 ifp->if_capabilities |= IFCAP_LINKSTATE; 435 ifp->if_capenable |= IFCAP_LINKSTATE; 436 437 ether_ifattach(ifp, eaddr); 438 *ret = ifp; 439 return 0; 440} 441/* unregister from the system and drop the final refcount */ 442void 443nm_vi_detach(struct ifnet *ifp) 444{ 445 nm_vi_free_index(((char *)IF_LLADDR(ifp))[5]); 446 ether_ifdetach(ifp); 447 if_free(ifp); 448} 449 450/* 451 * In order to track whether pages are still mapped, we hook into 452 * the standard cdev_pager and intercept the constructor and 453 * destructor. 454 */ 455 456struct netmap_vm_handle_t { 457 struct cdev *dev; 458 struct netmap_priv_d *priv; 459}; 460 461 462static int 463netmap_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, 464 vm_ooffset_t foff, struct ucred *cred, u_short *color) 465{ 466 struct netmap_vm_handle_t *vmh = handle; 467 468 if (netmap_verbose) 469 D("handle %p size %jd prot %d foff %jd", 470 handle, (intmax_t)size, prot, (intmax_t)foff); 471 if (color) 472 *color = 0; 473 dev_ref(vmh->dev); 474 return 0; 475} 476 477 478static void 479netmap_dev_pager_dtor(void *handle) 480{ 481 struct netmap_vm_handle_t *vmh = handle; 482 struct cdev *dev = vmh->dev; 483 struct netmap_priv_d *priv = vmh->priv; 484 485 if (netmap_verbose) 486 D("handle %p", handle); 487 netmap_dtor(priv); 488 free(vmh, M_DEVBUF); 489 dev_rel(dev); 490} 491 492 493static int 494netmap_dev_pager_fault(vm_object_t object, vm_ooffset_t offset, 495 int prot, vm_page_t *mres) 496{ 497 struct netmap_vm_handle_t *vmh = object->handle; 498 struct netmap_priv_d *priv = vmh->priv; 499 struct netmap_adapter *na = priv->np_na; 500 vm_paddr_t paddr; 501 vm_page_t page; 502 vm_memattr_t memattr; 503 vm_pindex_t pidx; 504 505 ND("object %p offset %jd prot %d mres %p", 506 object, (intmax_t)offset, prot, mres); 507 memattr = object->memattr; 508 pidx = OFF_TO_IDX(offset); 509 paddr = netmap_mem_ofstophys(na->nm_mem, offset); 510 if (paddr == 0) 511 return VM_PAGER_FAIL; 512 513 if (((*mres)->flags & PG_FICTITIOUS) != 0) { 514 /* 515 * If the passed in result page is a fake page, update it with 516 * the new physical address. 517 */ 518 page = *mres; 519 vm_page_updatefake(page, paddr, memattr); 520 } else { 521 /* 522 * Replace the passed in reqpage page with our own fake page and 523 * free up the all of the original pages. 524 */ 525#ifndef VM_OBJECT_WUNLOCK /* FreeBSD < 10.x */ 526#define VM_OBJECT_WUNLOCK VM_OBJECT_UNLOCK 527#define VM_OBJECT_WLOCK VM_OBJECT_LOCK 528#endif /* VM_OBJECT_WUNLOCK */ 529 530 VM_OBJECT_WUNLOCK(object); 531 page = vm_page_getfake(paddr, memattr); 532 VM_OBJECT_WLOCK(object); 533 vm_page_lock(*mres); 534 vm_page_free(*mres); 535 vm_page_unlock(*mres); 536 *mres = page; 537 vm_page_insert(page, object, pidx); 538 } 539 page->valid = VM_PAGE_BITS_ALL; 540 return (VM_PAGER_OK); 541} 542 543 544static struct cdev_pager_ops netmap_cdev_pager_ops = { 545 .cdev_pg_ctor = netmap_dev_pager_ctor, 546 .cdev_pg_dtor = netmap_dev_pager_dtor, 547 .cdev_pg_fault = netmap_dev_pager_fault, 548}; 549 550 551static int 552netmap_mmap_single(struct cdev *cdev, vm_ooffset_t *foff, 553 vm_size_t objsize, vm_object_t *objp, int prot) 554{ 555 int error; 556 struct netmap_vm_handle_t *vmh; 557 struct netmap_priv_d *priv; 558 vm_object_t obj; 559 560 if (netmap_verbose) 561 D("cdev %p foff %jd size %jd objp %p prot %d", cdev, 562 (intmax_t )*foff, (intmax_t )objsize, objp, prot); 563 564 vmh = malloc(sizeof(struct netmap_vm_handle_t), M_DEVBUF, 565 M_NOWAIT | M_ZERO); 566 if (vmh == NULL) 567 return ENOMEM; 568 vmh->dev = cdev; 569 570 NMG_LOCK(); 571 error = devfs_get_cdevpriv((void**)&priv); 572 if (error) 573 goto err_unlock; 574 if (priv->np_nifp == NULL) { 575 error = EINVAL; 576 goto err_unlock; 577 } 578 vmh->priv = priv; 579 priv->np_refs++; 580 NMG_UNLOCK(); 581 582 obj = cdev_pager_allocate(vmh, OBJT_DEVICE, 583 &netmap_cdev_pager_ops, objsize, prot, 584 *foff, NULL); 585 if (obj == NULL) { 586 D("cdev_pager_allocate failed"); 587 error = EINVAL; 588 goto err_deref; 589 } 590 591 *objp = obj; 592 return 0; 593 594err_deref: 595 NMG_LOCK(); 596 priv->np_refs--; 597err_unlock: 598 NMG_UNLOCK(); 599// err: 600 free(vmh, M_DEVBUF); 601 return error; 602} 603 604/* 605 * On FreeBSD the close routine is only called on the last close on 606 * the device (/dev/netmap) so we cannot do anything useful. 607 * To track close() on individual file descriptors we pass netmap_dtor() to 608 * devfs_set_cdevpriv() on open(). The FreeBSD kernel will call the destructor 609 * when the last fd pointing to the device is closed. 610 * 611 * Note that FreeBSD does not even munmap() on close() so we also have 612 * to track mmap() ourselves, and postpone the call to 613 * netmap_dtor() is called when the process has no open fds and no active 614 * memory maps on /dev/netmap, as in linux. 615 */ 616static int 617netmap_close(struct cdev *dev, int fflag, int devtype, struct thread *td) 618{ 619 if (netmap_verbose) 620 D("dev %p fflag 0x%x devtype %d td %p", 621 dev, fflag, devtype, td); 622 return 0; 623} 624 625 626static int 627netmap_open(struct cdev *dev, int oflags, int devtype, struct thread *td) 628{ 629 struct netmap_priv_d *priv; 630 int error; 631 632 (void)dev; 633 (void)oflags; 634 (void)devtype; 635 (void)td; 636 637 priv = malloc(sizeof(struct netmap_priv_d), M_DEVBUF, 638 M_NOWAIT | M_ZERO); 639 if (priv == NULL) 640 return ENOMEM; 641 priv->np_refs = 1; 642 error = devfs_set_cdevpriv(priv, netmap_dtor); 643 if (error) { 644 free(priv, M_DEVBUF); 645 } 646 return error; 647} 648 649/******************** kqueue support ****************/ 650 651/* 652 * The OS_selwakeup also needs to issue a KNOTE_UNLOCKED. 653 * We use a non-zero argument to distinguish the call from the one 654 * in kevent_scan() which instead also needs to run netmap_poll(). 655 * The knote uses a global mutex for the time being. We might 656 * try to reuse the one in the si, but it is not allocated 657 * permanently so it might be a bit tricky. 658 * 659 * The *kqfilter function registers one or another f_event 660 * depending on read or write mode. 661 * In the call to f_event() td_fpop is NULL so any child function 662 * calling devfs_get_cdevpriv() would fail - and we need it in 663 * netmap_poll(). As a workaround we store priv into kn->kn_hook 664 * and pass it as first argument to netmap_poll(), which then 665 * uses the failure to tell that we are called from f_event() 666 * and do not need the selrecord(). 667 */ 668 669 670void 671freebsd_selwakeup(struct nm_selinfo *si, int pri) 672{ 673 if (netmap_verbose) 674 D("on knote %p", &si->si.si_note); 675 selwakeuppri(&si->si, pri); 676 /* use a non-zero hint to tell the notification from the 677 * call done in kqueue_scan() which uses 0 678 */ 679 KNOTE_UNLOCKED(&si->si.si_note, 0x100 /* notification */); 680} 681 682static void 683netmap_knrdetach(struct knote *kn) 684{ 685 struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook; 686 struct selinfo *si = &priv->np_si[NR_RX]->si; 687 688 D("remove selinfo %p", si); 689 knlist_remove(&si->si_note, kn, 0); 690} 691 692static void 693netmap_knwdetach(struct knote *kn) 694{ 695 struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook; 696 struct selinfo *si = &priv->np_si[NR_TX]->si; 697 698 D("remove selinfo %p", si); 699 knlist_remove(&si->si_note, kn, 0); 700} 701 702/* 703 * callback from notifies (generated externally) and our 704 * calls to kevent(). The former we just return 1 (ready) 705 * since we do not know better. 706 * In the latter we call netmap_poll and return 0/1 accordingly. 707 */ 708static int 709netmap_knrw(struct knote *kn, long hint, int events) 710{ 711 struct netmap_priv_d *priv; 712 int revents; 713 714 if (hint != 0) { 715 ND(5, "call from notify"); 716 return 1; /* assume we are ready */ 717 } 718 priv = kn->kn_hook; 719 /* the notification may come from an external thread, 720 * in which case we do not want to run the netmap_poll 721 * This should be filtered above, but check just in case. 722 */ 723 if (curthread != priv->np_td) { /* should not happen */ 724 RD(5, "curthread changed %p %p", curthread, priv->np_td); 725 return 1; 726 } else { 727 revents = netmap_poll((void *)priv, events, curthread); 728 return (events & revents) ? 1 : 0; 729 } 730} 731 732static int 733netmap_knread(struct knote *kn, long hint) 734{ 735 return netmap_knrw(kn, hint, POLLIN); 736} 737 738static int 739netmap_knwrite(struct knote *kn, long hint) 740{ 741 return netmap_knrw(kn, hint, POLLOUT); 742} 743 744static struct filterops netmap_rfiltops = { 745 .f_isfd = 1, 746 .f_detach = netmap_knrdetach, 747 .f_event = netmap_knread, 748}; 749 750static struct filterops netmap_wfiltops = { 751 .f_isfd = 1, 752 .f_detach = netmap_knwdetach, 753 .f_event = netmap_knwrite, 754}; 755 756 757/* 758 * This is called when a thread invokes kevent() to record 759 * a change in the configuration of the kqueue(). 760 * The 'priv' should be the same as in the netmap device. 761 */ 762static int 763netmap_kqfilter(struct cdev *dev, struct knote *kn) 764{ 765 struct netmap_priv_d *priv; 766 int error; 767 struct netmap_adapter *na; 768 struct nm_selinfo *si; 769 int ev = kn->kn_filter; 770 771 if (ev != EVFILT_READ && ev != EVFILT_WRITE) { 772 D("bad filter request %d", ev); 773 return 1; 774 } 775 error = devfs_get_cdevpriv((void**)&priv); 776 if (error) { 777 D("device not yet setup"); 778 return 1; 779 } 780 na = priv->np_na; 781 if (na == NULL) { 782 D("no netmap adapter for this file descriptor"); 783 return 1; 784 } 785 /* the si is indicated in the priv */ 786 si = priv->np_si[(ev == EVFILT_WRITE) ? NR_TX : NR_RX]; 787 // XXX lock(priv) ? 788 kn->kn_fop = (ev == EVFILT_WRITE) ? 789 &netmap_wfiltops : &netmap_rfiltops; 790 kn->kn_hook = priv; 791 knlist_add(&si->si.si_note, kn, 1); 792 // XXX unlock(priv) 793 ND("register %p %s td %p priv %p kn %p np_nifp %p kn_fp/fpop %s", 794 na, na->ifp->if_xname, curthread, priv, kn, 795 priv->np_nifp, 796 kn->kn_fp == curthread->td_fpop ? "match" : "MISMATCH"); 797 return 0; 798} 799 800struct cdevsw netmap_cdevsw = { 801 .d_version = D_VERSION, 802 .d_name = "netmap", 803 .d_open = netmap_open, 804 .d_mmap_single = netmap_mmap_single, 805 .d_ioctl = netmap_ioctl, 806 .d_poll = netmap_poll, 807 .d_kqfilter = netmap_kqfilter, 808 .d_close = netmap_close, 809}; 810/*--- end of kqueue support ----*/ 811 812/* 813 * Kernel entry point. 814 * 815 * Initialize/finalize the module and return. 816 * 817 * Return 0 on success, errno on failure. 818 */ 819static int 820netmap_loader(__unused struct module *module, int event, __unused void *arg) 821{ 822 int error = 0; 823 824 switch (event) { 825 case MOD_LOAD: 826 error = netmap_init(); 827 break; 828 829 case MOD_UNLOAD: 830 netmap_fini(); 831 break; 832 833 default: 834 error = EOPNOTSUPP; 835 break; 836 } 837 838 return (error); 839} 840 841 842DEV_MODULE(netmap, netmap_loader, NULL); 843MODULE_VERSION(netmap, 1); 844