1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo 5 * Copyright (C) 2013-2016 Universita` di Pisa 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 /* 31 * $FreeBSD$ 32 * 33 * The header contains the definitions of constants and function 34 * prototypes used only in kernelspace. 35 */ 36 37 #ifndef _NET_NETMAP_KERN_H_ 38 #define _NET_NETMAP_KERN_H_ 39 40 #if defined(linux) 41 42 #if defined(CONFIG_NETMAP_EXTMEM) 43 #define WITH_EXTMEM 44 #endif 45 #if defined(CONFIG_NETMAP_VALE) 46 #define WITH_VALE 47 #endif 48 #if defined(CONFIG_NETMAP_PIPE) 49 #define WITH_PIPES 50 #endif 51 #if defined(CONFIG_NETMAP_MONITOR) 52 #define WITH_MONITOR 53 #endif 54 #if defined(CONFIG_NETMAP_GENERIC) 55 #define WITH_GENERIC 56 #endif 57 #if defined(CONFIG_NETMAP_PTNETMAP) 58 #define WITH_PTNETMAP 59 #endif 60 #if defined(CONFIG_NETMAP_SINK) 61 #define WITH_SINK 62 #endif 63 #if defined(CONFIG_NETMAP_NULL) 64 #define WITH_NMNULL 65 #endif 66 67 #elif defined (_WIN32) 68 #define WITH_VALE // comment out to disable VALE support 69 #define WITH_PIPES 70 #define WITH_MONITOR 71 #define WITH_GENERIC 72 #define WITH_NMNULL 73 74 #else /* neither linux nor windows */ 75 #define WITH_VALE // comment out to disable VALE support 76 #define WITH_PIPES 77 #define WITH_MONITOR 78 #define WITH_GENERIC 79 #define WITH_EXTMEM 80 #define WITH_NMNULL 81 #endif 82 83 #if defined(__FreeBSD__) 84 #include <sys/selinfo.h> 85 86 #define likely(x) __builtin_expect((long)!!(x), 1L) 87 #define unlikely(x) __builtin_expect((long)!!(x), 0L) 88 #define __user 89 90 #define NM_LOCK_T struct mtx /* low level spinlock, used to protect queues */ 91 92 #define NM_MTX_T struct sx /* OS-specific mutex (sleepable) */ 93 #define NM_MTX_INIT(m) sx_init(&(m), #m) 94 #define NM_MTX_DESTROY(m) sx_destroy(&(m)) 95 #define NM_MTX_LOCK(m) sx_xlock(&(m)) 96 #define NM_MTX_SPINLOCK(m) while (!sx_try_xlock(&(m))) ; 97 #define NM_MTX_UNLOCK(m) sx_xunlock(&(m)) 98 #define NM_MTX_ASSERT(m) sx_assert(&(m), SA_XLOCKED) 99 100 #define NM_SELINFO_T struct nm_selinfo 101 #define NM_SELRECORD_T struct thread 102 #define MBUF_LEN(m) ((m)->m_pkthdr.len) 103 #define MBUF_TXQ(m) ((m)->m_pkthdr.flowid) 104 #define MBUF_TRANSMIT(na, ifp, m) ((na)->if_transmit(ifp, m)) 105 #define GEN_TX_MBUF_IFP(m) ((m)->m_pkthdr.rcvif) 106 107 #define NM_ATOMIC_T volatile int /* required by atomic/bitops.h */ 108 /* atomic operations */ 109 #include <machine/atomic.h> 110 #define NM_ATOMIC_TEST_AND_SET(p) (!atomic_cmpset_acq_int((p), 0, 1)) 111 #define NM_ATOMIC_CLEAR(p) atomic_store_rel_int((p), 0) 112 113 #define WNA(_ifp) (_ifp)->if_netmap 114 115 struct netmap_adapter *netmap_getna(if_t ifp); 116 117 #define MBUF_REFCNT(m) ((m)->m_ext.ext_count) 118 #define SET_MBUF_REFCNT(m, x) (m)->m_ext.ext_count = x 119 120 #define MBUF_QUEUED(m) 1 121 122 struct nm_selinfo { 123 /* Support for select(2) and poll(2). */ 124 struct selinfo si; 125 /* Support for kqueue(9). See comments in netmap_freebsd.c */ 126 struct taskqueue *ntfytq; 127 struct task ntfytask; 128 struct mtx m; 129 char mtxname[32]; 130 int kqueue_users; 131 }; 132 133 134 struct hrtimer { 135 /* Not used in FreeBSD. */ 136 }; 137 138 #define NM_BNS_GET(b) 139 #define NM_BNS_PUT(b) 140 141 #elif defined (linux) 142 143 #define NM_LOCK_T safe_spinlock_t // see bsd_glue.h 144 #define NM_SELINFO_T wait_queue_head_t 145 #define MBUF_LEN(m) ((m)->len) 146 #define MBUF_TRANSMIT(na, ifp, m) \ 147 ({ \ 148 /* Avoid infinite recursion with generic. */ \ 149 m->priority = NM_MAGIC_PRIORITY_TX; \ 150 (((struct net_device_ops *)(na)->if_transmit)->ndo_start_xmit(m, ifp)); \ 151 0; \ 152 }) 153 154 /* See explanation in nm_os_generic_xmit_frame. */ 155 #define GEN_TX_MBUF_IFP(m) ((struct ifnet *)skb_shinfo(m)->destructor_arg) 156 157 #define NM_ATOMIC_T volatile long unsigned int 158 159 #define NM_MTX_T struct mutex /* OS-specific sleepable lock */ 160 #define NM_MTX_INIT(m) mutex_init(&(m)) 161 #define NM_MTX_DESTROY(m) do { (void)(m); } while (0) 162 #define NM_MTX_LOCK(m) mutex_lock(&(m)) 163 #define NM_MTX_UNLOCK(m) mutex_unlock(&(m)) 164 #define NM_MTX_ASSERT(m) mutex_is_locked(&(m)) 165 166 #ifndef DEV_NETMAP 167 #define DEV_NETMAP 168 #endif /* DEV_NETMAP */ 169 170 #elif defined (__APPLE__) 171 172 #warning apple support is incomplete. 173 #define likely(x) __builtin_expect(!!(x), 1) 174 #define unlikely(x) __builtin_expect(!!(x), 0) 175 #define NM_LOCK_T IOLock * 176 #define NM_SELINFO_T struct selinfo 177 #define MBUF_LEN(m) ((m)->m_pkthdr.len) 178 179 #elif defined (_WIN32) 180 #include "../../../WINDOWS/win_glue.h" 181 182 #define NM_SELRECORD_T IO_STACK_LOCATION 183 #define NM_SELINFO_T win_SELINFO // see win_glue.h 184 #define NM_LOCK_T win_spinlock_t // see win_glue.h 185 #define NM_MTX_T KGUARDED_MUTEX /* OS-specific mutex (sleepable) */ 186 187 #define NM_MTX_INIT(m) KeInitializeGuardedMutex(&m); 188 #define NM_MTX_DESTROY(m) do { (void)(m); } while (0) 189 #define NM_MTX_LOCK(m) KeAcquireGuardedMutex(&(m)) 190 #define NM_MTX_UNLOCK(m) KeReleaseGuardedMutex(&(m)) 191 #define NM_MTX_ASSERT(m) assert(&m.Count>0) 192 193 //These linknames are for the NDIS driver 194 #define NETMAP_NDIS_LINKNAME_STRING L"\\DosDevices\\NMAPNDIS" 195 #define NETMAP_NDIS_NTDEVICE_STRING L"\\Device\\NMAPNDIS" 196 197 //Definition of internal driver-to-driver ioctl codes 198 #define NETMAP_KERNEL_XCHANGE_POINTERS _IO('i', 180) 199 #define NETMAP_KERNEL_SEND_SHUTDOWN_SIGNAL _IO_direct('i', 195) 200 201 typedef struct hrtimer{ 202 KTIMER timer; 203 BOOLEAN active; 204 KDPC deferred_proc; 205 }; 206 207 /* MSVC does not have likely/unlikely support */ 208 #ifdef _MSC_VER 209 #define likely(x) (x) 210 #define unlikely(x) (x) 211 #else 212 #define likely(x) __builtin_expect((long)!!(x), 1L) 213 #define unlikely(x) __builtin_expect((long)!!(x), 0L) 214 #endif //_MSC_VER 215 216 #else 217 218 #error unsupported platform 219 220 #endif /* end - platform-specific code */ 221 222 #ifndef _WIN32 /* support for emulated sysctl */ 223 #define SYSBEGIN(x) 224 #define SYSEND 225 #endif /* _WIN32 */ 226 227 #define NM_ACCESS_ONCE(x) (*(volatile __typeof__(x) *)&(x)) 228 229 #define NMG_LOCK_T NM_MTX_T 230 #define NMG_LOCK_INIT() NM_MTX_INIT(netmap_global_lock) 231 #define NMG_LOCK_DESTROY() NM_MTX_DESTROY(netmap_global_lock) 232 #define NMG_LOCK() NM_MTX_LOCK(netmap_global_lock) 233 #define NMG_UNLOCK() NM_MTX_UNLOCK(netmap_global_lock) 234 #define NMG_LOCK_ASSERT() NM_MTX_ASSERT(netmap_global_lock) 235 236 #if defined(__FreeBSD__) 237 #define nm_prerr_int printf 238 #define nm_prinf_int printf 239 #elif defined (_WIN32) 240 #define nm_prerr_int DbgPrint 241 #define nm_prinf_int DbgPrint 242 #elif defined(linux) 243 #define nm_prerr_int(fmt, arg...) printk(KERN_ERR fmt, ##arg) 244 #define nm_prinf_int(fmt, arg...) printk(KERN_INFO fmt, ##arg) 245 #endif 246 247 #define nm_prinf(format, ...) \ 248 do { \ 249 struct timeval __xxts; \ 250 microtime(&__xxts); \ 251 nm_prinf_int("%03d.%06d [%4d] %-25s " format "\n",\ 252 (int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec, \ 253 __LINE__, __FUNCTION__, ##__VA_ARGS__); \ 254 } while (0) 255 256 #define nm_prerr(format, ...) \ 257 do { \ 258 struct timeval __xxts; \ 259 microtime(&__xxts); \ 260 nm_prerr_int("%03d.%06d [%4d] %-25s " format "\n",\ 261 (int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec, \ 262 __LINE__, __FUNCTION__, ##__VA_ARGS__); \ 263 } while (0) 264 265 /* Disabled printf (used to be nm_prdis). */ 266 #define nm_prdis(format, ...) 267 268 /* Rate limited, lps indicates how many per second. */ 269 #define nm_prlim(lps, format, ...) \ 270 do { \ 271 static int t0, __cnt; \ 272 if (t0 != time_second) { \ 273 t0 = time_second; \ 274 __cnt = 0; \ 275 } \ 276 if (__cnt++ < lps) \ 277 nm_prinf(format, ##__VA_ARGS__); \ 278 } while (0) 279 280 struct netmap_adapter; 281 struct nm_bdg_fwd; 282 struct nm_bridge; 283 struct netmap_priv_d; 284 struct nm_bdg_args; 285 286 /* os-specific NM_SELINFO_T initialization/destruction functions */ 287 int nm_os_selinfo_init(NM_SELINFO_T *, const char *name); 288 void nm_os_selinfo_uninit(NM_SELINFO_T *); 289 290 const char *nm_dump_buf(char *p, int len, int lim, char *dst); 291 292 void nm_os_selwakeup(NM_SELINFO_T *si); 293 void nm_os_selrecord(NM_SELRECORD_T *sr, NM_SELINFO_T *si); 294 295 int nm_os_ifnet_init(void); 296 void nm_os_ifnet_fini(void); 297 void nm_os_ifnet_lock(void); 298 void nm_os_ifnet_unlock(void); 299 300 unsigned nm_os_ifnet_mtu(struct ifnet *ifp); 301 302 void nm_os_get_module(void); 303 void nm_os_put_module(void); 304 305 void netmap_make_zombie(struct ifnet *); 306 void netmap_undo_zombie(struct ifnet *); 307 308 /* os independent alloc/realloc/free */ 309 void *nm_os_malloc(size_t); 310 void *nm_os_vmalloc(size_t); 311 void *nm_os_realloc(void *, size_t new_size, size_t old_size); 312 void nm_os_free(void *); 313 void nm_os_vfree(void *); 314 315 /* os specific attach/detach enter/exit-netmap-mode routines */ 316 void nm_os_onattach(struct ifnet *); 317 void nm_os_ondetach(struct ifnet *); 318 void nm_os_onenter(struct ifnet *); 319 void nm_os_onexit(struct ifnet *); 320 321 /* passes a packet up to the host stack. 322 * If the packet is sent (or dropped) immediately it returns NULL, 323 * otherwise it links the packet to prev and returns m. 324 * In this case, a final call with m=NULL and prev != NULL will send up 325 * the entire chain to the host stack. 326 */ 327 void *nm_os_send_up(struct ifnet *, struct mbuf *m, struct mbuf *prev); 328 329 int nm_os_mbuf_has_seg_offld(struct mbuf *m); 330 int nm_os_mbuf_has_csum_offld(struct mbuf *m); 331 332 #include "netmap_mbq.h" 333 334 extern NMG_LOCK_T netmap_global_lock; 335 336 enum txrx { NR_RX = 0, NR_TX = 1, NR_TXRX }; 337 338 static __inline const char* 339 nm_txrx2str(enum txrx t) 340 { 341 return (t== NR_RX ? "RX" : "TX"); 342 } 343 344 static __inline enum txrx 345 nm_txrx_swap(enum txrx t) 346 { 347 return (t== NR_RX ? NR_TX : NR_RX); 348 } 349 350 #define for_rx_tx(t) for ((t) = 0; (t) < NR_TXRX; (t)++) 351 352 #ifdef WITH_MONITOR 353 struct netmap_zmon_list { 354 struct netmap_kring *next; 355 struct netmap_kring *prev; 356 }; 357 #endif /* WITH_MONITOR */ 358 359 /* 360 * private, kernel view of a ring. Keeps track of the status of 361 * a ring across system calls. 362 * 363 * nr_hwcur index of the next buffer to refill. 364 * It corresponds to ring->head 365 * at the time the system call returns. 366 * 367 * nr_hwtail index of the first buffer owned by the kernel. 368 * On RX, hwcur->hwtail are receive buffers 369 * not yet released. hwcur is advanced following 370 * ring->head, hwtail is advanced on incoming packets, 371 * and a wakeup is generated when hwtail passes ring->cur 372 * On TX, hwcur->rcur have been filled by the sender 373 * but not sent yet to the NIC; rcur->hwtail are available 374 * for new transmissions, and hwtail->hwcur-1 are pending 375 * transmissions not yet acknowledged. 376 * 377 * The indexes in the NIC and netmap rings are offset by nkr_hwofs slots. 378 * This is so that, on a reset, buffers owned by userspace are not 379 * modified by the kernel. In particular: 380 * RX rings: the next empty buffer (hwtail + hwofs) coincides with 381 * the next empty buffer as known by the hardware (next_to_check or so). 382 * TX rings: hwcur + hwofs coincides with next_to_send 383 * 384 * The following fields are used to implement lock-free copy of packets 385 * from input to output ports in VALE switch: 386 * nkr_hwlease buffer after the last one being copied. 387 * A writer in nm_bdg_flush reserves N buffers 388 * from nr_hwlease, advances it, then does the 389 * copy outside the lock. 390 * In RX rings (used for VALE ports), 391 * nkr_hwtail <= nkr_hwlease < nkr_hwcur+N-1 392 * In TX rings (used for NIC or host stack ports) 393 * nkr_hwcur <= nkr_hwlease < nkr_hwtail 394 * nkr_leases array of nkr_num_slots where writers can report 395 * completion of their block. NR_NOSLOT (~0) indicates 396 * that the writer has not finished yet 397 * nkr_lease_idx index of next free slot in nr_leases, to be assigned 398 * 399 * The kring is manipulated by txsync/rxsync and generic netmap function. 400 * 401 * Concurrent rxsync or txsync on the same ring are prevented through 402 * by nm_kr_(try)lock() which in turn uses nr_busy. This is all we need 403 * for NIC rings, and for TX rings attached to the host stack. 404 * 405 * RX rings attached to the host stack use an mbq (rx_queue) on both 406 * rxsync_from_host() and netmap_transmit(). The mbq is protected 407 * by its internal lock. 408 * 409 * RX rings attached to the VALE switch are accessed by both senders 410 * and receiver. They are protected through the q_lock on the RX ring. 411 */ 412 struct netmap_kring { 413 struct netmap_ring *ring; 414 415 uint32_t nr_hwcur; /* should be nr_hwhead */ 416 uint32_t nr_hwtail; 417 418 /* 419 * Copies of values in user rings, so we do not need to look 420 * at the ring (which could be modified). These are set in the 421 * *sync_prologue()/finalize() routines. 422 */ 423 uint32_t rhead; 424 uint32_t rcur; 425 uint32_t rtail; 426 427 uint32_t nr_kflags; /* private driver flags */ 428 #define NKR_PENDINTR 0x1 // Pending interrupt. 429 #define NKR_EXCLUSIVE 0x2 /* exclusive binding */ 430 #define NKR_FORWARD 0x4 /* (host ring only) there are 431 packets to forward 432 */ 433 #define NKR_NEEDRING 0x8 /* ring needed even if users==0 434 * (used internally by pipes and 435 * by ptnetmap host ports) 436 */ 437 #define NKR_NOINTR 0x10 /* don't use interrupts on this ring */ 438 #define NKR_FAKERING 0x20 /* don't allocate/free buffers */ 439 440 uint32_t nr_mode; 441 uint32_t nr_pending_mode; 442 #define NKR_NETMAP_OFF 0x0 443 #define NKR_NETMAP_ON 0x1 444 445 uint32_t nkr_num_slots; 446 447 /* 448 * On a NIC reset, the NIC ring indexes may be reset but the 449 * indexes in the netmap rings remain the same. nkr_hwofs 450 * keeps track of the offset between the two. 451 * 452 * Moreover, during reset, we can restore only the subset of 453 * the NIC ring that corresponds to the kernel-owned part of 454 * the netmap ring. The rest of the slots must be restored 455 * by the *sync routines when the user releases more slots. 456 * The nkr_to_refill field keeps track of the number of slots 457 * that still need to be restored. 458 */ 459 int32_t nkr_hwofs; 460 int32_t nkr_to_refill; 461 462 /* last_reclaim is opaque marker to help reduce the frequency 463 * of operations such as reclaiming tx buffers. A possible use 464 * is set it to ticks and do the reclaim only once per tick. 465 */ 466 uint64_t last_reclaim; 467 468 469 NM_SELINFO_T si; /* poll/select wait queue */ 470 NM_LOCK_T q_lock; /* protects kring and ring. */ 471 NM_ATOMIC_T nr_busy; /* prevent concurrent syscalls */ 472 473 /* the adapter the owns this kring */ 474 struct netmap_adapter *na; 475 476 /* the adapter that wants to be notified when this kring has 477 * new slots available. This is usually the same as the above, 478 * but wrappers may let it point to themselves 479 */ 480 struct netmap_adapter *notify_na; 481 482 /* The following fields are for VALE switch support */ 483 struct nm_bdg_fwd *nkr_ft; 484 uint32_t *nkr_leases; 485 #define NR_NOSLOT ((uint32_t)~0) /* used in nkr_*lease* */ 486 uint32_t nkr_hwlease; 487 uint32_t nkr_lease_idx; 488 489 /* while nkr_stopped is set, no new [tr]xsync operations can 490 * be started on this kring. 491 * This is used by netmap_disable_all_rings() 492 * to find a synchronization point where critical data 493 * structures pointed to by the kring can be added or removed 494 */ 495 volatile int nkr_stopped; 496 497 /* Support for adapters without native netmap support. 498 * On tx rings we preallocate an array of tx buffers 499 * (same size as the netmap ring), on rx rings we 500 * store incoming mbufs in a queue that is drained by 501 * a rxsync. 502 */ 503 struct mbuf **tx_pool; 504 struct mbuf *tx_event; /* TX event used as a notification */ 505 NM_LOCK_T tx_event_lock; /* protects the tx_event mbuf */ 506 struct mbq rx_queue; /* intercepted rx mbufs. */ 507 508 uint32_t users; /* existing bindings for this ring */ 509 510 uint32_t ring_id; /* kring identifier */ 511 enum txrx tx; /* kind of ring (tx or rx) */ 512 char name[64]; /* diagnostic */ 513 514 /* [tx]sync callback for this kring. 515 * The default nm_kring_create callback (netmap_krings_create) 516 * sets the nm_sync callback of each hardware tx(rx) kring to 517 * the corresponding nm_txsync(nm_rxsync) taken from the 518 * netmap_adapter; moreover, it sets the sync callback 519 * of the host tx(rx) ring to netmap_txsync_to_host 520 * (netmap_rxsync_from_host). 521 * 522 * Overrides: the above configuration is not changed by 523 * any of the nm_krings_create callbacks. 524 */ 525 int (*nm_sync)(struct netmap_kring *kring, int flags); 526 int (*nm_notify)(struct netmap_kring *kring, int flags); 527 528 #ifdef WITH_PIPES 529 struct netmap_kring *pipe; /* if this is a pipe ring, 530 * pointer to the other end 531 */ 532 uint32_t pipe_tail; /* hwtail updated by the other end */ 533 #endif /* WITH_PIPES */ 534 535 /* mask for the offset-related part of the ptr field in the slots */ 536 uint64_t offset_mask; 537 /* maximum user-specified offset, as stipulated at bind time. 538 * Larger offset requests will be silently capped to offset_max. 539 */ 540 uint64_t offset_max; 541 /* minimum gap between two consecutive offsets into the same 542 * buffer, as stipulated at bind time. This is used to choose 543 * the hwbuf_len, but is not otherwise checked for compliance 544 * at runtime. 545 */ 546 uint64_t offset_gap; 547 548 /* size of hardware buffer. This may be less than the size of 549 * the netmap buffers because of non-zero offsets, or because 550 * the netmap buffer size exceeds the capability of the hardware. 551 */ 552 uint64_t hwbuf_len; 553 554 /* required alignment (in bytes) for the buffers used by this ring. 555 * Netmap buffers are aligned to cachelines, which should suffice 556 * for most NICs. If the user is passing offsets, though, we need 557 * to check that the resulting buf address complies with any 558 * alignment restriction. 559 */ 560 uint64_t buf_align; 561 562 /* hardware specific logic for the selection of the hwbuf_len */ 563 int (*nm_bufcfg)(struct netmap_kring *kring, uint64_t target); 564 565 int (*save_notify)(struct netmap_kring *kring, int flags); 566 567 #ifdef WITH_MONITOR 568 /* array of krings that are monitoring this kring */ 569 struct netmap_kring **monitors; 570 uint32_t max_monitors; /* current size of the monitors array */ 571 uint32_t n_monitors; /* next unused entry in the monitor array */ 572 uint32_t mon_pos[NR_TXRX]; /* index of this ring in the monitored ring array */ 573 uint32_t mon_tail; /* last seen slot on rx */ 574 575 /* circular list of zero-copy monitors */ 576 struct netmap_zmon_list zmon_list[NR_TXRX]; 577 578 /* 579 * Monitors work by intercepting the sync and notify callbacks of the 580 * monitored krings. This is implemented by replacing the pointers 581 * above and saving the previous ones in mon_* pointers below 582 */ 583 int (*mon_sync)(struct netmap_kring *kring, int flags); 584 int (*mon_notify)(struct netmap_kring *kring, int flags); 585 586 #endif 587 } 588 #ifdef _WIN32 589 __declspec(align(64)); 590 #else 591 __attribute__((__aligned__(64))); 592 #endif 593 594 /* return 1 iff the kring needs to be turned on */ 595 static inline int 596 nm_kring_pending_on(struct netmap_kring *kring) 597 { 598 return kring->nr_pending_mode == NKR_NETMAP_ON && 599 kring->nr_mode == NKR_NETMAP_OFF; 600 } 601 602 /* return 1 iff the kring needs to be turned off */ 603 static inline int 604 nm_kring_pending_off(struct netmap_kring *kring) 605 { 606 return kring->nr_pending_mode == NKR_NETMAP_OFF && 607 kring->nr_mode == NKR_NETMAP_ON; 608 } 609 610 /* return the next index, with wraparound */ 611 static inline uint32_t 612 nm_next(uint32_t i, uint32_t lim) 613 { 614 return unlikely (i == lim) ? 0 : i + 1; 615 } 616 617 618 /* return the previous index, with wraparound */ 619 static inline uint32_t 620 nm_prev(uint32_t i, uint32_t lim) 621 { 622 return unlikely (i == 0) ? lim : i - 1; 623 } 624 625 626 /* 627 * 628 * Here is the layout for the Rx and Tx rings. 629 630 RxRING TxRING 631 632 +-----------------+ +-----------------+ 633 | | | | 634 | free | | free | 635 +-----------------+ +-----------------+ 636 head->| owned by user |<-hwcur | not sent to nic |<-hwcur 637 | | | yet | 638 +-----------------+ | | 639 cur->| available to | | | 640 | user, not read | +-----------------+ 641 | yet | cur->| (being | 642 | | | prepared) | 643 | | | | 644 +-----------------+ + ------ + 645 tail->| |<-hwtail | |<-hwlease 646 | (being | ... | | ... 647 | prepared) | ... | | ... 648 +-----------------+ ... | | ... 649 | |<-hwlease +-----------------+ 650 | | tail->| |<-hwtail 651 | | | | 652 | | | | 653 | | | | 654 +-----------------+ +-----------------+ 655 656 * The cur/tail (user view) and hwcur/hwtail (kernel view) 657 * are used in the normal operation of the card. 658 * 659 * When a ring is the output of a switch port (Rx ring for 660 * a VALE port, Tx ring for the host stack or NIC), slots 661 * are reserved in blocks through 'hwlease' which points 662 * to the next unused slot. 663 * On an Rx ring, hwlease is always after hwtail, 664 * and completions cause hwtail to advance. 665 * On a Tx ring, hwlease is always between cur and hwtail, 666 * and completions cause cur to advance. 667 * 668 * nm_kr_space() returns the maximum number of slots that 669 * can be assigned. 670 * nm_kr_lease() reserves the required number of buffers, 671 * advances nkr_hwlease and also returns an entry in 672 * a circular array where completions should be reported. 673 */ 674 675 struct lut_entry; 676 #ifdef __FreeBSD__ 677 #define plut_entry lut_entry 678 #endif 679 680 struct netmap_lut { 681 struct lut_entry *lut; 682 struct plut_entry *plut; 683 uint32_t objtotal; /* max buffer index */ 684 uint32_t objsize; /* buffer size */ 685 }; 686 687 struct netmap_vp_adapter; // forward 688 struct nm_bridge; 689 690 /* Struct to be filled by nm_config callbacks. */ 691 struct nm_config_info { 692 unsigned num_tx_rings; 693 unsigned num_rx_rings; 694 unsigned num_tx_descs; 695 unsigned num_rx_descs; 696 unsigned rx_buf_maxsize; 697 }; 698 699 /* 700 * default type for the magic field. 701 * May be overridden in glue code. 702 */ 703 #ifndef NM_OS_MAGIC 704 #define NM_OS_MAGIC uint32_t 705 #endif /* !NM_OS_MAGIC */ 706 707 /* 708 * The "struct netmap_adapter" extends the "struct adapter" 709 * (or equivalent) device descriptor. 710 * It contains all base fields needed to support netmap operation. 711 * There are in fact different types of netmap adapters 712 * (native, generic, VALE switch...) so a netmap_adapter is 713 * just the first field in the derived type. 714 */ 715 struct netmap_adapter { 716 /* 717 * On linux we do not have a good way to tell if an interface 718 * is netmap-capable. So we always use the following trick: 719 * NA(ifp) points here, and the first entry (which hopefully 720 * always exists and is at least 32 bits) contains a magic 721 * value which we can use to detect that the interface is good. 722 */ 723 NM_OS_MAGIC magic; 724 uint32_t na_flags; /* enabled, and other flags */ 725 #define NAF_SKIP_INTR 1 /* use the regular interrupt handler. 726 * useful during initialization 727 */ 728 #define NAF_SW_ONLY 2 /* forward packets only to sw adapter */ 729 #define NAF_BDG_MAYSLEEP 4 /* the bridge is allowed to sleep when 730 * forwarding packets coming from this 731 * interface 732 */ 733 #define NAF_MEM_OWNER 8 /* the adapter uses its own memory area 734 * that cannot be changed 735 */ 736 #define NAF_NATIVE 16 /* the adapter is native. 737 * Virtual ports (non persistent vale ports, 738 * pipes, monitors...) should never use 739 * this flag. 740 */ 741 #define NAF_NETMAP_ON 32 /* netmap is active (either native or 742 * emulated). Where possible (e.g. FreeBSD) 743 * IFCAP_NETMAP also mirrors this flag. 744 */ 745 #define NAF_HOST_RINGS 64 /* the adapter supports the host rings */ 746 #define NAF_FORCE_NATIVE 128 /* the adapter is always NATIVE */ 747 /* free */ 748 #define NAF_MOREFRAG 512 /* the adapter supports NS_MOREFRAG */ 749 #define NAF_OFFSETS 1024 /* the adapter supports the slot offsets */ 750 #define NAF_HOST_ALL 2048 /* the adapter wants as many host rings as hw */ 751 #define NAF_ZOMBIE (1U<<30) /* the nic driver has been unloaded */ 752 #define NAF_BUSY (1U<<31) /* the adapter is used internally and 753 * cannot be registered from userspace 754 */ 755 int active_fds; /* number of user-space descriptors using this 756 interface, which is equal to the number of 757 struct netmap_if objs in the mapped region. */ 758 759 u_int num_rx_rings; /* number of adapter receive rings */ 760 u_int num_tx_rings; /* number of adapter transmit rings */ 761 u_int num_host_rx_rings; /* number of host receive rings */ 762 u_int num_host_tx_rings; /* number of host transmit rings */ 763 764 u_int num_tx_desc; /* number of descriptor in each queue */ 765 u_int num_rx_desc; 766 767 /* tx_rings and rx_rings are private but allocated as a 768 * contiguous chunk of memory. Each array has N+K entries, 769 * N for the hardware rings and K for the host rings. 770 */ 771 struct netmap_kring **tx_rings; /* array of TX rings. */ 772 struct netmap_kring **rx_rings; /* array of RX rings. */ 773 774 void *tailroom; /* space below the rings array */ 775 /* (used for leases) */ 776 777 778 NM_SELINFO_T si[NR_TXRX]; /* global wait queues */ 779 780 /* count users of the global wait queues */ 781 int si_users[NR_TXRX]; 782 783 void *pdev; /* used to store pci device */ 784 785 /* copy of if_qflush and if_transmit pointers, to intercept 786 * packets from the network stack when netmap is active. 787 */ 788 int (*if_transmit)(struct ifnet *, struct mbuf *); 789 790 /* copy of if_input for netmap_send_up() */ 791 void (*if_input)(struct ifnet *, struct mbuf *); 792 793 /* Back reference to the parent ifnet struct. Used for 794 * hardware ports (emulated netmap included). */ 795 struct ifnet *ifp; /* adapter is ifp->if_softc */ 796 797 /*---- callbacks for this netmap adapter -----*/ 798 /* 799 * nm_dtor() is the cleanup routine called when destroying 800 * the adapter. 801 * Called with NMG_LOCK held. 802 * 803 * nm_register() is called on NIOCREGIF and close() to enter 804 * or exit netmap mode on the NIC 805 * Called with NNG_LOCK held. 806 * 807 * nm_txsync() pushes packets to the underlying hw/switch 808 * 809 * nm_rxsync() collects packets from the underlying hw/switch 810 * 811 * nm_config() returns configuration information from the OS 812 * Called with NMG_LOCK held. 813 * 814 * nm_bufcfg() 815 * the purpose of this callback is to fill the kring->hwbuf_len 816 * (l) and kring->buf_align fields. The l value is most important 817 * for RX rings, where we want to disallow writes outside of the 818 * netmap buffer. The l value must be computed taking into account 819 * the stipulated max_offset (o), possibly increased if there are 820 * alignment constraints, the maxframe (m), if known, and the 821 * current NETMAP_BUF_SIZE (b) of the memory region used by the 822 * adapter. We want the largest supported l such that o + l <= b. 823 * If m is known to be <= b - o, the callback may also choose the 824 * largest l <= m, ignoring the offset. The buf_align field is 825 * most important for TX rings when there are offsets. The user 826 * will see this value in the ring->buf_align field. Misaligned 827 * offsets will cause the corresponding packets to be silently 828 * dropped. 829 * 830 * nm_krings_create() create and init the tx_rings and 831 * rx_rings arrays of kring structures. In particular, 832 * set the nm_sync callbacks for each ring. 833 * There is no need to also allocate the corresponding 834 * netmap_rings, since netmap_mem_rings_create() will always 835 * be called to provide the missing ones. 836 * Called with NNG_LOCK held. 837 * 838 * nm_krings_delete() cleanup and delete the tx_rings and rx_rings 839 * arrays 840 * Called with NMG_LOCK held. 841 * 842 * nm_notify() is used to act after data have become available 843 * (or the stopped state of the ring has changed) 844 * For hw devices this is typically a selwakeup(), 845 * but for NIC/host ports attached to a switch (or vice-versa) 846 * we also need to invoke the 'txsync' code downstream. 847 * This callback pointer is actually used only to initialize 848 * kring->nm_notify. 849 * Return values are the same as for netmap_rx_irq(). 850 */ 851 void (*nm_dtor)(struct netmap_adapter *); 852 853 int (*nm_register)(struct netmap_adapter *, int onoff); 854 void (*nm_intr)(struct netmap_adapter *, int onoff); 855 856 int (*nm_txsync)(struct netmap_kring *kring, int flags); 857 int (*nm_rxsync)(struct netmap_kring *kring, int flags); 858 int (*nm_notify)(struct netmap_kring *kring, int flags); 859 int (*nm_bufcfg)(struct netmap_kring *kring, uint64_t target); 860 #define NAF_FORCE_READ 1 861 #define NAF_FORCE_RECLAIM 2 862 #define NAF_CAN_FORWARD_DOWN 4 863 /* return configuration information */ 864 int (*nm_config)(struct netmap_adapter *, struct nm_config_info *info); 865 int (*nm_krings_create)(struct netmap_adapter *); 866 void (*nm_krings_delete)(struct netmap_adapter *); 867 /* 868 * nm_bdg_attach() initializes the na_vp field to point 869 * to an adapter that can be attached to a VALE switch. If the 870 * current adapter is already a VALE port, na_vp is simply a cast; 871 * otherwise, na_vp points to a netmap_bwrap_adapter. 872 * If applicable, this callback also initializes na_hostvp, 873 * that can be used to connect the adapter host rings to the 874 * switch. 875 * Called with NMG_LOCK held. 876 * 877 * nm_bdg_ctl() is called on the actual attach/detach to/from 878 * to/from the switch, to perform adapter-specific 879 * initializations 880 * Called with NMG_LOCK held. 881 */ 882 int (*nm_bdg_attach)(const char *bdg_name, struct netmap_adapter *, 883 struct nm_bridge *); 884 int (*nm_bdg_ctl)(struct nmreq_header *, struct netmap_adapter *); 885 886 /* adapter used to attach this adapter to a VALE switch (if any) */ 887 struct netmap_vp_adapter *na_vp; 888 /* adapter used to attach the host rings of this adapter 889 * to a VALE switch (if any) */ 890 struct netmap_vp_adapter *na_hostvp; 891 892 /* standard refcount to control the lifetime of the adapter 893 * (it should be equal to the lifetime of the corresponding ifp) 894 */ 895 int na_refcount; 896 897 /* memory allocator (opaque) 898 * We also cache a pointer to the lut_entry for translating 899 * buffer addresses, the total number of buffers and the buffer size. 900 */ 901 struct netmap_mem_d *nm_mem; 902 struct netmap_mem_d *nm_mem_prev; 903 struct netmap_lut na_lut; 904 905 /* additional information attached to this adapter 906 * by other netmap subsystems. Currently used by 907 * bwrap, LINUX/v1000 and ptnetmap 908 */ 909 void *na_private; 910 911 /* array of pipes that have this adapter as a parent */ 912 struct netmap_pipe_adapter **na_pipes; 913 int na_next_pipe; /* next free slot in the array */ 914 int na_max_pipes; /* size of the array */ 915 916 /* Offset of ethernet header for each packet. */ 917 u_int virt_hdr_len; 918 919 /* Max number of bytes that the NIC can store in the buffer 920 * referenced by each RX descriptor. This translates to the maximum 921 * bytes that a single netmap slot can reference. Larger packets 922 * require NS_MOREFRAG support. */ 923 unsigned rx_buf_maxsize; 924 925 char name[NETMAP_REQ_IFNAMSIZ]; /* used at least by pipes */ 926 927 #ifdef WITH_MONITOR 928 unsigned long monitor_id; /* debugging */ 929 #endif 930 }; 931 932 static __inline u_int 933 nma_get_ndesc(struct netmap_adapter *na, enum txrx t) 934 { 935 return (t == NR_TX ? na->num_tx_desc : na->num_rx_desc); 936 } 937 938 static __inline void 939 nma_set_ndesc(struct netmap_adapter *na, enum txrx t, u_int v) 940 { 941 if (t == NR_TX) 942 na->num_tx_desc = v; 943 else 944 na->num_rx_desc = v; 945 } 946 947 static __inline u_int 948 nma_get_nrings(struct netmap_adapter *na, enum txrx t) 949 { 950 return (t == NR_TX ? na->num_tx_rings : na->num_rx_rings); 951 } 952 953 static __inline u_int 954 nma_get_host_nrings(struct netmap_adapter *na, enum txrx t) 955 { 956 return (t == NR_TX ? na->num_host_tx_rings : na->num_host_rx_rings); 957 } 958 959 static __inline void 960 nma_set_nrings(struct netmap_adapter *na, enum txrx t, u_int v) 961 { 962 if (t == NR_TX) 963 na->num_tx_rings = v; 964 else 965 na->num_rx_rings = v; 966 } 967 968 static __inline void 969 nma_set_host_nrings(struct netmap_adapter *na, enum txrx t, u_int v) 970 { 971 if (t == NR_TX) 972 na->num_host_tx_rings = v; 973 else 974 na->num_host_rx_rings = v; 975 } 976 977 static __inline struct netmap_kring** 978 NMR(struct netmap_adapter *na, enum txrx t) 979 { 980 return (t == NR_TX ? na->tx_rings : na->rx_rings); 981 } 982 983 int nma_intr_enable(struct netmap_adapter *na, int onoff); 984 985 /* 986 * If the NIC is owned by the kernel 987 * (i.e., bridge), neither another bridge nor user can use it; 988 * if the NIC is owned by a user, only users can share it. 989 * Evaluation must be done under NMG_LOCK(). 990 */ 991 #define NETMAP_OWNED_BY_KERN(na) ((na)->na_flags & NAF_BUSY) 992 #define NETMAP_OWNED_BY_ANY(na) \ 993 (NETMAP_OWNED_BY_KERN(na) || ((na)->active_fds > 0)) 994 995 /* 996 * derived netmap adapters for various types of ports 997 */ 998 struct netmap_vp_adapter { /* VALE software port */ 999 struct netmap_adapter up; 1000 1001 /* 1002 * Bridge support: 1003 * 1004 * bdg_port is the port number used in the bridge; 1005 * na_bdg points to the bridge this NA is attached to. 1006 */ 1007 int bdg_port; 1008 struct nm_bridge *na_bdg; 1009 int retry; 1010 int autodelete; /* remove the ifp on last reference */ 1011 1012 /* Maximum Frame Size, used in bdg_mismatch_datapath() */ 1013 u_int mfs; 1014 /* Last source MAC on this port */ 1015 uint64_t last_smac; 1016 }; 1017 1018 1019 struct netmap_hw_adapter { /* physical device */ 1020 struct netmap_adapter up; 1021 1022 #ifdef linux 1023 struct net_device_ops nm_ndo; 1024 struct ethtool_ops nm_eto; 1025 #endif 1026 const struct ethtool_ops* save_ethtool; 1027 1028 int (*nm_hw_register)(struct netmap_adapter *, int onoff); 1029 }; 1030 1031 #ifdef WITH_GENERIC 1032 /* Mitigation support. */ 1033 struct nm_generic_mit { 1034 struct hrtimer mit_timer; 1035 int mit_pending; 1036 int mit_ring_idx; /* index of the ring being mitigated */ 1037 struct netmap_adapter *mit_na; /* backpointer */ 1038 }; 1039 1040 struct netmap_generic_adapter { /* emulated device */ 1041 struct netmap_hw_adapter up; 1042 1043 /* Pointer to a previously used netmap adapter. */ 1044 struct netmap_adapter *prev; 1045 1046 /* Emulated netmap adapters support: 1047 * - save_if_input saves the if_input hook (FreeBSD); 1048 * - mit implements rx interrupt mitigation; 1049 */ 1050 void (*save_if_input)(struct ifnet *, struct mbuf *); 1051 1052 struct nm_generic_mit *mit; 1053 #ifdef linux 1054 netdev_tx_t (*save_start_xmit)(struct mbuf *, struct ifnet *); 1055 #endif 1056 /* Is the adapter able to use multiple RX slots to scatter 1057 * each packet pushed up by the driver? */ 1058 int rxsg; 1059 1060 /* Is the transmission path controlled by a netmap-aware 1061 * device queue (i.e. qdisc on linux)? */ 1062 int txqdisc; 1063 }; 1064 #endif /* WITH_GENERIC */ 1065 1066 static __inline u_int 1067 netmap_real_rings(struct netmap_adapter *na, enum txrx t) 1068 { 1069 return nma_get_nrings(na, t) + 1070 !!(na->na_flags & NAF_HOST_RINGS) * nma_get_host_nrings(na, t); 1071 } 1072 1073 /* account for fake rings */ 1074 static __inline u_int 1075 netmap_all_rings(struct netmap_adapter *na, enum txrx t) 1076 { 1077 return max(nma_get_nrings(na, t) + 1, netmap_real_rings(na, t)); 1078 } 1079 1080 int netmap_default_bdg_attach(const char *name, struct netmap_adapter *na, 1081 struct nm_bridge *); 1082 struct nm_bdg_polling_state; 1083 /* 1084 * Bridge wrapper for non VALE ports attached to a VALE switch. 1085 * 1086 * The real device must already have its own netmap adapter (hwna). 1087 * The bridge wrapper and the hwna adapter share the same set of 1088 * netmap rings and buffers, but they have two separate sets of 1089 * krings descriptors, with tx/rx meanings swapped: 1090 * 1091 * netmap 1092 * bwrap krings rings krings hwna 1093 * +------+ +------+ +-----+ +------+ +------+ 1094 * |tx_rings->| |\ /| |----| |<-tx_rings| 1095 * | | +------+ \ / +-----+ +------+ | | 1096 * | | X | | 1097 * | | / \ | | 1098 * | | +------+/ \+-----+ +------+ | | 1099 * |rx_rings->| | | |----| |<-rx_rings| 1100 * | | +------+ +-----+ +------+ | | 1101 * +------+ +------+ 1102 * 1103 * - packets coming from the bridge go to the brwap rx rings, 1104 * which are also the hwna tx rings. The bwrap notify callback 1105 * will then complete the hwna tx (see netmap_bwrap_notify). 1106 * 1107 * - packets coming from the outside go to the hwna rx rings, 1108 * which are also the bwrap tx rings. The (overwritten) hwna 1109 * notify method will then complete the bridge tx 1110 * (see netmap_bwrap_intr_notify). 1111 * 1112 * The bridge wrapper may optionally connect the hwna 'host' rings 1113 * to the bridge. This is done by using a second port in the 1114 * bridge and connecting it to the 'host' netmap_vp_adapter 1115 * contained in the netmap_bwrap_adapter. The brwap host adapter 1116 * cross-links the hwna host rings in the same way as shown above. 1117 * 1118 * - packets coming from the bridge and directed to the host stack 1119 * are handled by the bwrap host notify callback 1120 * (see netmap_bwrap_host_notify) 1121 * 1122 * - packets coming from the host stack are still handled by the 1123 * overwritten hwna notify callback (netmap_bwrap_intr_notify), 1124 * but are diverted to the host adapter depending on the ring number. 1125 * 1126 */ 1127 struct netmap_bwrap_adapter { 1128 struct netmap_vp_adapter up; 1129 struct netmap_vp_adapter host; /* for host rings */ 1130 struct netmap_adapter *hwna; /* the underlying device */ 1131 1132 /* 1133 * When we attach a physical interface to the bridge, we 1134 * allow the controlling process to terminate, so we need 1135 * a place to store the n_detmap_priv_d data structure. 1136 * This is only done when physical interfaces 1137 * are attached to a bridge. 1138 */ 1139 struct netmap_priv_d *na_kpriv; 1140 struct nm_bdg_polling_state *na_polling_state; 1141 /* we overwrite the hwna->na_vp pointer, so we save 1142 * here its original value, to be restored at detach 1143 */ 1144 struct netmap_vp_adapter *saved_na_vp; 1145 int (*nm_intr_notify)(struct netmap_kring *kring, int flags); 1146 }; 1147 int nm_is_bwrap(struct netmap_adapter *na); 1148 int nm_bdg_polling(struct nmreq_header *hdr); 1149 1150 int netmap_bdg_attach(struct nmreq_header *hdr, void *auth_token); 1151 int netmap_bdg_detach(struct nmreq_header *hdr, void *auth_token); 1152 #ifdef WITH_VALE 1153 int netmap_vale_list(struct nmreq_header *hdr); 1154 int netmap_vi_create(struct nmreq_header *hdr, int); 1155 int nm_vi_create(struct nmreq_header *); 1156 int nm_vi_destroy(const char *name); 1157 #else /* !WITH_VALE */ 1158 #define netmap_vi_create(hdr, a) (EOPNOTSUPP) 1159 #endif /* WITH_VALE */ 1160 1161 #ifdef WITH_PIPES 1162 1163 #define NM_MAXPIPES 64 /* max number of pipes per adapter */ 1164 1165 struct netmap_pipe_adapter { 1166 /* pipe identifier is up.name */ 1167 struct netmap_adapter up; 1168 1169 #define NM_PIPE_ROLE_MASTER 0x1 1170 #define NM_PIPE_ROLE_SLAVE 0x2 1171 int role; /* either NM_PIPE_ROLE_MASTER or NM_PIPE_ROLE_SLAVE */ 1172 1173 struct netmap_adapter *parent; /* adapter that owns the memory */ 1174 struct netmap_pipe_adapter *peer; /* the other end of the pipe */ 1175 int peer_ref; /* 1 iff we are holding a ref to the peer */ 1176 struct ifnet *parent_ifp; /* maybe null */ 1177 1178 u_int parent_slot; /* index in the parent pipe array */ 1179 }; 1180 1181 #endif /* WITH_PIPES */ 1182 1183 #ifdef WITH_NMNULL 1184 struct netmap_null_adapter { 1185 struct netmap_adapter up; 1186 }; 1187 #endif /* WITH_NMNULL */ 1188 1189 1190 /* return slots reserved to rx clients; used in drivers */ 1191 static inline uint32_t 1192 nm_kr_rxspace(struct netmap_kring *k) 1193 { 1194 int space = k->nr_hwtail - k->nr_hwcur; 1195 if (space < 0) 1196 space += k->nkr_num_slots; 1197 nm_prdis("preserving %d rx slots %d -> %d", space, k->nr_hwcur, k->nr_hwtail); 1198 1199 return space; 1200 } 1201 1202 /* return slots reserved to tx clients */ 1203 #define nm_kr_txspace(_k) nm_kr_rxspace(_k) 1204 1205 1206 /* True if no space in the tx ring, only valid after txsync_prologue */ 1207 static inline int 1208 nm_kr_txempty(struct netmap_kring *kring) 1209 { 1210 return kring->rhead == kring->nr_hwtail; 1211 } 1212 1213 /* True if no more completed slots in the rx ring, only valid after 1214 * rxsync_prologue */ 1215 #define nm_kr_rxempty(_k) nm_kr_txempty(_k) 1216 1217 /* True if the application needs to wait for more space on the ring 1218 * (more received packets or more free tx slots). 1219 * Only valid after *xsync_prologue. */ 1220 static inline int 1221 nm_kr_wouldblock(struct netmap_kring *kring) 1222 { 1223 return kring->rcur == kring->nr_hwtail; 1224 } 1225 1226 /* 1227 * protect against multiple threads using the same ring. 1228 * also check that the ring has not been stopped or locked 1229 */ 1230 #define NM_KR_BUSY 1 /* some other thread is syncing the ring */ 1231 #define NM_KR_STOPPED 2 /* unbounded stop (ifconfig down or driver unload) */ 1232 #define NM_KR_LOCKED 3 /* bounded, brief stop for mutual exclusion */ 1233 1234 1235 /* release the previously acquired right to use the *sync() methods of the ring */ 1236 static __inline void nm_kr_put(struct netmap_kring *kr) 1237 { 1238 NM_ATOMIC_CLEAR(&kr->nr_busy); 1239 } 1240 1241 1242 /* true if the ifp that backed the adapter has disappeared (e.g., the 1243 * driver has been unloaded) 1244 */ 1245 static inline int nm_iszombie(struct netmap_adapter *na); 1246 1247 /* try to obtain exclusive right to issue the *sync() operations on the ring. 1248 * The right is obtained and must be later relinquished via nm_kr_put() if and 1249 * only if nm_kr_tryget() returns 0. 1250 * If can_sleep is 1 there are only two other possible outcomes: 1251 * - the function returns NM_KR_BUSY 1252 * - the function returns NM_KR_STOPPED and sets the POLLERR bit in *perr 1253 * (if non-null) 1254 * In both cases the caller will typically skip the ring, possibly collecting 1255 * errors along the way. 1256 * If the calling context does not allow sleeping, the caller must pass 0 in can_sleep. 1257 * In the latter case, the function may also return NM_KR_LOCKED and leave *perr 1258 * untouched: ideally, the caller should try again at a later time. 1259 */ 1260 static __inline int nm_kr_tryget(struct netmap_kring *kr, int can_sleep, int *perr) 1261 { 1262 int busy = 1, stopped; 1263 /* check a first time without taking the lock 1264 * to avoid starvation for nm_kr_get() 1265 */ 1266 retry: 1267 stopped = kr->nkr_stopped; 1268 if (unlikely(stopped)) { 1269 goto stop; 1270 } 1271 busy = NM_ATOMIC_TEST_AND_SET(&kr->nr_busy); 1272 /* we should not return NM_KR_BUSY if the ring was 1273 * actually stopped, so check another time after 1274 * the barrier provided by the atomic operation 1275 */ 1276 stopped = kr->nkr_stopped; 1277 if (unlikely(stopped)) { 1278 goto stop; 1279 } 1280 1281 if (unlikely(nm_iszombie(kr->na))) { 1282 stopped = NM_KR_STOPPED; 1283 goto stop; 1284 } 1285 1286 return unlikely(busy) ? NM_KR_BUSY : 0; 1287 1288 stop: 1289 if (!busy) 1290 nm_kr_put(kr); 1291 if (stopped == NM_KR_STOPPED) { 1292 /* if POLLERR is defined we want to use it to simplify netmap_poll(). 1293 * Otherwise, any non-zero value will do. 1294 */ 1295 #ifdef POLLERR 1296 #define NM_POLLERR POLLERR 1297 #else 1298 #define NM_POLLERR 1 1299 #endif /* POLLERR */ 1300 if (perr) 1301 *perr |= NM_POLLERR; 1302 #undef NM_POLLERR 1303 } else if (can_sleep) { 1304 tsleep(kr, 0, "NM_KR_TRYGET", 4); 1305 goto retry; 1306 } 1307 return stopped; 1308 } 1309 1310 /* put the ring in the 'stopped' state and wait for the current user (if any) to 1311 * notice. stopped must be either NM_KR_STOPPED or NM_KR_LOCKED 1312 */ 1313 static __inline void nm_kr_stop(struct netmap_kring *kr, int stopped) 1314 { 1315 kr->nkr_stopped = stopped; 1316 while (NM_ATOMIC_TEST_AND_SET(&kr->nr_busy)) 1317 tsleep(kr, 0, "NM_KR_GET", 4); 1318 } 1319 1320 /* restart a ring after a stop */ 1321 static __inline void nm_kr_start(struct netmap_kring *kr) 1322 { 1323 kr->nkr_stopped = 0; 1324 nm_kr_put(kr); 1325 } 1326 1327 1328 /* 1329 * The following functions are used by individual drivers to 1330 * support netmap operation. 1331 * 1332 * netmap_attach() initializes a struct netmap_adapter, allocating the 1333 * struct netmap_ring's and the struct selinfo. 1334 * 1335 * netmap_detach() frees the memory allocated by netmap_attach(). 1336 * 1337 * netmap_transmit() replaces the if_transmit routine of the interface, 1338 * and is used to intercept packets coming from the stack. 1339 * 1340 * netmap_load_map/netmap_reload_map are helper routines to set/reset 1341 * the dmamap for a packet buffer 1342 * 1343 * netmap_reset() is a helper routine to be called in the hw driver 1344 * when reinitializing a ring. It should not be called by 1345 * virtual ports (vale, pipes, monitor) 1346 */ 1347 int netmap_attach(struct netmap_adapter *); 1348 int netmap_attach_ext(struct netmap_adapter *, size_t size, int override_reg); 1349 void netmap_detach(struct ifnet *); 1350 int netmap_transmit(struct ifnet *, struct mbuf *); 1351 struct netmap_slot *netmap_reset(struct netmap_adapter *na, 1352 enum txrx tx, u_int n, u_int new_cur); 1353 int netmap_ring_reinit(struct netmap_kring *); 1354 int netmap_rings_config_get(struct netmap_adapter *, struct nm_config_info *); 1355 1356 /* Return codes for netmap_*x_irq. */ 1357 enum { 1358 /* Driver should do normal interrupt processing, e.g. because 1359 * the interface is not in netmap mode. */ 1360 NM_IRQ_PASS = 0, 1361 /* Port is in netmap mode, and the interrupt work has been 1362 * completed. The driver does not have to notify netmap 1363 * again before the next interrupt. */ 1364 NM_IRQ_COMPLETED = -1, 1365 /* Port is in netmap mode, but the interrupt work has not been 1366 * completed. The driver has to make sure netmap will be 1367 * notified again soon, even if no more interrupts come (e.g. 1368 * on Linux the driver should not call napi_complete()). */ 1369 NM_IRQ_RESCHED = -2, 1370 }; 1371 1372 /* default functions to handle rx/tx interrupts */ 1373 int netmap_rx_irq(struct ifnet *, u_int, u_int *); 1374 #define netmap_tx_irq(_n, _q) netmap_rx_irq(_n, _q, NULL) 1375 int netmap_common_irq(struct netmap_adapter *, u_int, u_int *work_done); 1376 1377 1378 #ifdef WITH_VALE 1379 /* functions used by external modules to interface with VALE */ 1380 #define netmap_vp_to_ifp(_vp) ((_vp)->up.ifp) 1381 #define netmap_ifp_to_vp(_ifp) (NA(_ifp)->na_vp) 1382 #define netmap_ifp_to_host_vp(_ifp) (NA(_ifp)->na_hostvp) 1383 #define netmap_bdg_idx(_vp) ((_vp)->bdg_port) 1384 const char *netmap_bdg_name(struct netmap_vp_adapter *); 1385 #else /* !WITH_VALE */ 1386 #define netmap_vp_to_ifp(_vp) NULL 1387 #define netmap_ifp_to_vp(_ifp) NULL 1388 #define netmap_ifp_to_host_vp(_ifp) NULL 1389 #define netmap_bdg_idx(_vp) -1 1390 #endif /* WITH_VALE */ 1391 1392 static inline int 1393 nm_netmap_on(struct netmap_adapter *na) 1394 { 1395 return na && na->na_flags & NAF_NETMAP_ON; 1396 } 1397 1398 static inline int 1399 nm_native_on(struct netmap_adapter *na) 1400 { 1401 return nm_netmap_on(na) && (na->na_flags & NAF_NATIVE); 1402 } 1403 1404 static inline struct netmap_kring * 1405 netmap_kring_on(struct netmap_adapter *na, u_int q, enum txrx t) 1406 { 1407 struct netmap_kring *kring = NULL; 1408 1409 if (!nm_native_on(na)) 1410 return NULL; 1411 1412 if (t == NR_RX && q < na->num_rx_rings) 1413 kring = na->rx_rings[q]; 1414 else if (t == NR_TX && q < na->num_tx_rings) 1415 kring = na->tx_rings[q]; 1416 else 1417 return NULL; 1418 1419 return (kring->nr_mode == NKR_NETMAP_ON) ? kring : NULL; 1420 } 1421 1422 static inline int 1423 nm_iszombie(struct netmap_adapter *na) 1424 { 1425 return na == NULL || (na->na_flags & NAF_ZOMBIE); 1426 } 1427 1428 void nm_set_native_flags(struct netmap_adapter *); 1429 void nm_clear_native_flags(struct netmap_adapter *); 1430 1431 void netmap_krings_mode_commit(struct netmap_adapter *na, int onoff); 1432 1433 /* 1434 * nm_*sync_prologue() functions are used in ioctl/poll and ptnetmap 1435 * kthreads. 1436 * We need netmap_ring* parameter, because in ptnetmap it is decoupled 1437 * from host kring. 1438 * The user-space ring pointers (head/cur/tail) are shared through 1439 * CSB between host and guest. 1440 */ 1441 1442 /* 1443 * validates parameters in the ring/kring, returns a value for head 1444 * If any error, returns ring_size to force a reinit. 1445 */ 1446 uint32_t nm_txsync_prologue(struct netmap_kring *, struct netmap_ring *); 1447 1448 1449 /* 1450 * validates parameters in the ring/kring, returns a value for head 1451 * If any error, returns ring_size lim to force a reinit. 1452 */ 1453 uint32_t nm_rxsync_prologue(struct netmap_kring *, struct netmap_ring *); 1454 1455 1456 /* check/fix address and len in tx rings */ 1457 #if 1 /* debug version */ 1458 #define NM_CHECK_ADDR_LEN(_na, _a, _l) do { \ 1459 if (_a == NETMAP_BUF_BASE(_na) || _l > NETMAP_BUF_SIZE(_na)) { \ 1460 nm_prlim(5, "bad addr/len ring %d slot %d idx %d len %d", \ 1461 kring->ring_id, nm_i, slot->buf_idx, len); \ 1462 if (_l > NETMAP_BUF_SIZE(_na)) \ 1463 _l = NETMAP_BUF_SIZE(_na); \ 1464 } } while (0) 1465 #else /* no debug version */ 1466 #define NM_CHECK_ADDR_LEN(_na, _a, _l) do { \ 1467 if (_l > NETMAP_BUF_SIZE(_na)) \ 1468 _l = NETMAP_BUF_SIZE(_na); \ 1469 } while (0) 1470 #endif 1471 1472 #define NM_CHECK_ADDR_LEN_OFF(na_, l_, o_) do { \ 1473 if ((l_) + (o_) < (l_) || \ 1474 (l_) + (o_) > NETMAP_BUF_SIZE(na_)) { \ 1475 (l_) = NETMAP_BUF_SIZE(na_) - (o_); \ 1476 } } while (0) 1477 1478 1479 /*---------------------------------------------------------------*/ 1480 /* 1481 * Support routines used by netmap subsystems 1482 * (native drivers, VALE, generic, pipes, monitors, ...) 1483 */ 1484 1485 1486 /* common routine for all functions that create a netmap adapter. It performs 1487 * two main tasks: 1488 * - if the na points to an ifp, mark the ifp as netmap capable 1489 * using na as its native adapter; 1490 * - provide defaults for the setup callbacks and the memory allocator 1491 */ 1492 int netmap_attach_common(struct netmap_adapter *); 1493 /* fill priv->np_[tr]xq{first,last} using the ringid and flags information 1494 * coming from a struct nmreq_register 1495 */ 1496 int netmap_interp_ringid(struct netmap_priv_d *priv, struct nmreq_header *hdr); 1497 /* update the ring parameters (number and size of tx and rx rings). 1498 * It calls the nm_config callback, if available. 1499 */ 1500 int netmap_update_config(struct netmap_adapter *na); 1501 /* create and initialize the common fields of the krings array. 1502 * using the information that must be already available in the na. 1503 * tailroom can be used to request the allocation of additional 1504 * tailroom bytes after the krings array. This is used by 1505 * netmap_vp_adapter's (i.e., VALE ports) to make room for 1506 * leasing-related data structures 1507 */ 1508 int netmap_krings_create(struct netmap_adapter *na, u_int tailroom); 1509 /* deletes the kring array of the adapter. The array must have 1510 * been created using netmap_krings_create 1511 */ 1512 void netmap_krings_delete(struct netmap_adapter *na); 1513 1514 int netmap_hw_krings_create(struct netmap_adapter *na); 1515 void netmap_hw_krings_delete(struct netmap_adapter *na); 1516 1517 /* set the stopped/enabled status of ring 1518 * When stopping, they also wait for all current activity on the ring to 1519 * terminate. The status change is then notified using the na nm_notify 1520 * callback. 1521 */ 1522 void netmap_set_ring(struct netmap_adapter *, u_int ring_id, enum txrx, int stopped); 1523 /* set the stopped/enabled status of all rings of the adapter. */ 1524 void netmap_set_all_rings(struct netmap_adapter *, int stopped); 1525 /* convenience wrappers for netmap_set_all_rings */ 1526 void netmap_disable_all_rings(struct ifnet *); 1527 void netmap_enable_all_rings(struct ifnet *); 1528 1529 int netmap_buf_size_validate(const struct netmap_adapter *na, unsigned mtu); 1530 int netmap_do_regif(struct netmap_priv_d *priv, struct netmap_adapter *na, 1531 struct nmreq_header *); 1532 void netmap_do_unregif(struct netmap_priv_d *priv); 1533 1534 u_int nm_bound_var(u_int *v, u_int dflt, u_int lo, u_int hi, const char *msg); 1535 int netmap_get_na(struct nmreq_header *hdr, struct netmap_adapter **na, 1536 struct ifnet **ifp, struct netmap_mem_d *nmd, int create); 1537 void netmap_unget_na(struct netmap_adapter *na, struct ifnet *ifp); 1538 int netmap_get_hw_na(struct ifnet *ifp, 1539 struct netmap_mem_d *nmd, struct netmap_adapter **na); 1540 void netmap_mem_restore(struct netmap_adapter *na); 1541 1542 #ifdef WITH_VALE 1543 uint32_t netmap_vale_learning(struct nm_bdg_fwd *ft, uint8_t *dst_ring, 1544 struct netmap_vp_adapter *, void *private_data); 1545 1546 /* these are redefined in case of no VALE support */ 1547 int netmap_get_vale_na(struct nmreq_header *hdr, struct netmap_adapter **na, 1548 struct netmap_mem_d *nmd, int create); 1549 void *netmap_vale_create(const char *bdg_name, int *return_status); 1550 int netmap_vale_destroy(const char *bdg_name, void *auth_token); 1551 1552 extern unsigned int vale_max_bridges; 1553 1554 #else /* !WITH_VALE */ 1555 #define netmap_bdg_learning(_1, _2, _3, _4) 0 1556 #define netmap_get_vale_na(_1, _2, _3, _4) 0 1557 #define netmap_bdg_create(_1, _2) NULL 1558 #define netmap_bdg_destroy(_1, _2) 0 1559 #define vale_max_bridges 1 1560 #endif /* !WITH_VALE */ 1561 1562 #ifdef WITH_PIPES 1563 /* max number of pipes per device */ 1564 #define NM_MAXPIPES 64 /* XXX this should probably be a sysctl */ 1565 void netmap_pipe_dealloc(struct netmap_adapter *); 1566 int netmap_get_pipe_na(struct nmreq_header *hdr, struct netmap_adapter **na, 1567 struct netmap_mem_d *nmd, int create); 1568 #else /* !WITH_PIPES */ 1569 #define NM_MAXPIPES 0 1570 #define netmap_pipe_alloc(_1, _2) 0 1571 #define netmap_pipe_dealloc(_1) 1572 #define netmap_get_pipe_na(hdr, _2, _3, _4) \ 1573 ((strchr(hdr->nr_name, '{') != NULL || strchr(hdr->nr_name, '}') != NULL) ? EOPNOTSUPP : 0) 1574 #endif 1575 1576 #ifdef WITH_MONITOR 1577 int netmap_get_monitor_na(struct nmreq_header *hdr, struct netmap_adapter **na, 1578 struct netmap_mem_d *nmd, int create); 1579 void netmap_monitor_stop(struct netmap_adapter *na); 1580 #else 1581 #define netmap_get_monitor_na(hdr, _2, _3, _4) \ 1582 (((struct nmreq_register *)(uintptr_t)hdr->nr_body)->nr_flags & (NR_MONITOR_TX | NR_MONITOR_RX) ? EOPNOTSUPP : 0) 1583 #endif 1584 1585 #ifdef WITH_NMNULL 1586 int netmap_get_null_na(struct nmreq_header *hdr, struct netmap_adapter **na, 1587 struct netmap_mem_d *nmd, int create); 1588 #else /* !WITH_NMNULL */ 1589 #define netmap_get_null_na(hdr, _2, _3, _4) \ 1590 (((struct nmreq_register *)(uintptr_t)hdr->nr_body)->nr_flags & (NR_MONITOR_TX | NR_MONITOR_RX) ? EOPNOTSUPP : 0) 1591 #endif /* WITH_NMNULL */ 1592 1593 #ifdef CONFIG_NET_NS 1594 struct net *netmap_bns_get(void); 1595 void netmap_bns_put(struct net *); 1596 void netmap_bns_getbridges(struct nm_bridge **, u_int *); 1597 #else 1598 extern struct nm_bridge *nm_bridges; 1599 #define netmap_bns_get() 1600 #define netmap_bns_put(_1) 1601 #define netmap_bns_getbridges(b, n) \ 1602 do { *b = nm_bridges; *n = vale_max_bridges; } while (0) 1603 #endif 1604 1605 /* Various prototypes */ 1606 int netmap_poll(struct netmap_priv_d *, int events, NM_SELRECORD_T *td); 1607 int netmap_init(void); 1608 void netmap_fini(void); 1609 int netmap_get_memory(struct netmap_priv_d* p); 1610 void netmap_dtor(void *data); 1611 1612 int netmap_ioctl(struct netmap_priv_d *priv, u_long cmd, caddr_t data, 1613 struct thread *, int nr_body_is_user); 1614 int netmap_ioctl_legacy(struct netmap_priv_d *priv, u_long cmd, caddr_t data, 1615 struct thread *td); 1616 size_t nmreq_size_by_type(uint16_t nr_reqtype); 1617 1618 /* netmap_adapter creation/destruction */ 1619 1620 // #define NM_DEBUG_PUTGET 1 1621 1622 #ifdef NM_DEBUG_PUTGET 1623 1624 #define NM_DBG(f) __##f 1625 1626 void __netmap_adapter_get(struct netmap_adapter *na); 1627 1628 #define netmap_adapter_get(na) \ 1629 do { \ 1630 struct netmap_adapter *__na = na; \ 1631 __netmap_adapter_get(__na); \ 1632 nm_prinf("getting %p:%s -> %d", __na, (__na)->name, (__na)->na_refcount); \ 1633 } while (0) 1634 1635 int __netmap_adapter_put(struct netmap_adapter *na); 1636 1637 #define netmap_adapter_put(na) \ 1638 ({ \ 1639 struct netmap_adapter *__na = na; \ 1640 if (__na == NULL) \ 1641 nm_prinf("putting NULL"); \ 1642 else \ 1643 nm_prinf("putting %p:%s -> %d", __na, (__na)->name, (__na)->na_refcount - 1); \ 1644 __netmap_adapter_put(__na); \ 1645 }) 1646 1647 #else /* !NM_DEBUG_PUTGET */ 1648 1649 #define NM_DBG(f) f 1650 void netmap_adapter_get(struct netmap_adapter *na); 1651 int netmap_adapter_put(struct netmap_adapter *na); 1652 1653 #endif /* !NM_DEBUG_PUTGET */ 1654 1655 1656 /* 1657 * module variables 1658 */ 1659 #define NETMAP_BUF_BASE(_na) ((_na)->na_lut.lut[0].vaddr) 1660 #define NETMAP_BUF_SIZE(_na) ((_na)->na_lut.objsize) 1661 extern int netmap_no_pendintr; 1662 extern int netmap_verbose; 1663 #ifdef CONFIG_NETMAP_DEBUG 1664 extern int netmap_debug; /* for debugging */ 1665 #else /* !CONFIG_NETMAP_DEBUG */ 1666 #define netmap_debug (0) 1667 #endif /* !CONFIG_NETMAP_DEBUG */ 1668 enum { /* debug flags */ 1669 NM_DEBUG_ON = 1, /* generic debug messages */ 1670 NM_DEBUG_HOST = 0x2, /* debug host stack */ 1671 NM_DEBUG_RXSYNC = 0x10, /* debug on rxsync/txsync */ 1672 NM_DEBUG_TXSYNC = 0x20, 1673 NM_DEBUG_RXINTR = 0x100, /* debug on rx/tx intr (driver) */ 1674 NM_DEBUG_TXINTR = 0x200, 1675 NM_DEBUG_NIC_RXSYNC = 0x1000, /* debug on rx/tx intr (driver) */ 1676 NM_DEBUG_NIC_TXSYNC = 0x2000, 1677 NM_DEBUG_MEM = 0x4000, /* verbose memory allocations/deallocations */ 1678 NM_DEBUG_VALE = 0x8000, /* debug messages from memory allocators */ 1679 NM_DEBUG_BDG = NM_DEBUG_VALE, 1680 }; 1681 1682 extern int netmap_txsync_retry; 1683 extern int netmap_generic_hwcsum; 1684 extern int netmap_generic_mit; 1685 extern int netmap_generic_ringsize; 1686 extern int netmap_generic_rings; 1687 #ifdef linux 1688 extern int netmap_generic_txqdisc; 1689 #endif 1690 1691 /* 1692 * NA returns a pointer to the struct netmap adapter from the ifp. 1693 * WNA is os-specific and must be defined in glue code. 1694 */ 1695 #define NA(_ifp) ((struct netmap_adapter *)WNA(_ifp)) 1696 1697 /* 1698 * we provide a default implementation of NM_ATTACH_NA/NM_DETACH_NA 1699 * based on the WNA field. 1700 * Glue code may override this by defining its own NM_ATTACH_NA 1701 */ 1702 #ifndef NM_ATTACH_NA 1703 /* 1704 * On old versions of FreeBSD, NA(ifp) is a pspare. On linux we 1705 * overload another pointer in the netdev. 1706 * 1707 * We check if NA(ifp) is set and its first element has a related 1708 * magic value. The capenable is within the struct netmap_adapter. 1709 */ 1710 #define NETMAP_MAGIC 0x52697a7a 1711 1712 #define NM_NA_VALID(ifp) (NA(ifp) && \ 1713 ((uint32_t)(uintptr_t)NA(ifp) ^ NA(ifp)->magic) == NETMAP_MAGIC ) 1714 1715 #define NM_ATTACH_NA(ifp, na) do { \ 1716 WNA(ifp) = na; \ 1717 if (NA(ifp)) \ 1718 NA(ifp)->magic = \ 1719 ((uint32_t)(uintptr_t)NA(ifp)) ^ NETMAP_MAGIC; \ 1720 } while(0) 1721 #define NM_RESTORE_NA(ifp, na) WNA(ifp) = na; 1722 1723 #define NM_DETACH_NA(ifp) do { WNA(ifp) = NULL; } while (0) 1724 #define NM_NA_CLASH(ifp) (NA(ifp) && !NM_NA_VALID(ifp)) 1725 #endif /* !NM_ATTACH_NA */ 1726 1727 1728 #define NM_IS_NATIVE(ifp) (NM_NA_VALID(ifp) && NA(ifp)->nm_dtor == netmap_hw_dtor) 1729 1730 #if defined(__FreeBSD__) 1731 1732 /* Assigns the device IOMMU domain to an allocator. 1733 * Returns -ENOMEM in case the domain is different */ 1734 #define nm_iommu_group_id(dev) (-1) 1735 1736 /* Callback invoked by the dma machinery after a successful dmamap_load */ 1737 static void netmap_dmamap_cb(__unused void *arg, 1738 __unused bus_dma_segment_t * segs, __unused int nseg, __unused int error) 1739 { 1740 } 1741 1742 /* bus_dmamap_load wrapper: call aforementioned function if map != NULL. 1743 * XXX can we do it without a callback ? 1744 */ 1745 static inline int 1746 netmap_load_map(struct netmap_adapter *na, 1747 bus_dma_tag_t tag, bus_dmamap_t map, void *buf) 1748 { 1749 if (map) 1750 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE(na), 1751 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); 1752 return 0; 1753 } 1754 1755 static inline void 1756 netmap_unload_map(struct netmap_adapter *na, 1757 bus_dma_tag_t tag, bus_dmamap_t map) 1758 { 1759 if (map) 1760 bus_dmamap_unload(tag, map); 1761 } 1762 1763 #define netmap_sync_map(na, tag, map, sz, t) 1764 1765 /* update the map when a buffer changes. */ 1766 static inline void 1767 netmap_reload_map(struct netmap_adapter *na, 1768 bus_dma_tag_t tag, bus_dmamap_t map, void *buf) 1769 { 1770 if (map) { 1771 bus_dmamap_unload(tag, map); 1772 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE(na), 1773 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); 1774 } 1775 } 1776 1777 #elif defined(_WIN32) 1778 1779 #else /* linux */ 1780 1781 int nm_iommu_group_id(bus_dma_tag_t dev); 1782 #include <linux/dma-mapping.h> 1783 1784 /* 1785 * on linux we need 1786 * dma_map_single(&pdev->dev, virt_addr, len, direction) 1787 * dma_unmap_single(&adapter->pdev->dev, phys_addr, len, direction) 1788 */ 1789 #if 0 1790 struct e1000_buffer *buffer_info = &tx_ring->buffer_info[l]; 1791 /* set time_stamp *before* dma to help avoid a possible race */ 1792 buffer_info->time_stamp = jiffies; 1793 buffer_info->mapped_as_page = false; 1794 buffer_info->length = len; 1795 //buffer_info->next_to_watch = l; 1796 /* reload dma map */ 1797 dma_unmap_single(&adapter->pdev->dev, buffer_info->dma, 1798 NETMAP_BUF_SIZE, DMA_TO_DEVICE); 1799 buffer_info->dma = dma_map_single(&adapter->pdev->dev, 1800 addr, NETMAP_BUF_SIZE, DMA_TO_DEVICE); 1801 1802 if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) { 1803 nm_prerr("dma mapping error"); 1804 /* goto dma_error; See e1000_put_txbuf() */ 1805 /* XXX reset */ 1806 } 1807 tx_desc->buffer_addr = htole64(buffer_info->dma); //XXX 1808 1809 #endif 1810 1811 static inline int 1812 netmap_load_map(struct netmap_adapter *na, 1813 bus_dma_tag_t tag, bus_dmamap_t map, void *buf, u_int size) 1814 { 1815 if (map) { 1816 *map = dma_map_single(na->pdev, buf, size, 1817 DMA_BIDIRECTIONAL); 1818 if (dma_mapping_error(na->pdev, *map)) { 1819 *map = 0; 1820 return ENOMEM; 1821 } 1822 } 1823 return 0; 1824 } 1825 1826 static inline void 1827 netmap_unload_map(struct netmap_adapter *na, 1828 bus_dma_tag_t tag, bus_dmamap_t map, u_int sz) 1829 { 1830 if (*map) { 1831 dma_unmap_single(na->pdev, *map, sz, 1832 DMA_BIDIRECTIONAL); 1833 } 1834 } 1835 1836 #ifdef NETMAP_LINUX_HAVE_DMASYNC 1837 static inline void 1838 netmap_sync_map_cpu(struct netmap_adapter *na, 1839 bus_dma_tag_t tag, bus_dmamap_t map, u_int sz, enum txrx t) 1840 { 1841 if (*map) { 1842 dma_sync_single_for_cpu(na->pdev, *map, sz, 1843 (t == NR_TX ? DMA_TO_DEVICE : DMA_FROM_DEVICE)); 1844 } 1845 } 1846 1847 static inline void 1848 netmap_sync_map_dev(struct netmap_adapter *na, 1849 bus_dma_tag_t tag, bus_dmamap_t map, u_int sz, enum txrx t) 1850 { 1851 if (*map) { 1852 dma_sync_single_for_device(na->pdev, *map, sz, 1853 (t == NR_TX ? DMA_TO_DEVICE : DMA_FROM_DEVICE)); 1854 } 1855 } 1856 1857 static inline void 1858 netmap_reload_map(struct netmap_adapter *na, 1859 bus_dma_tag_t tag, bus_dmamap_t map, void *buf) 1860 { 1861 u_int sz = NETMAP_BUF_SIZE(na); 1862 1863 if (*map) { 1864 dma_unmap_single(na->pdev, *map, sz, 1865 DMA_BIDIRECTIONAL); 1866 } 1867 1868 *map = dma_map_single(na->pdev, buf, sz, 1869 DMA_BIDIRECTIONAL); 1870 } 1871 #else /* !NETMAP_LINUX_HAVE_DMASYNC */ 1872 #define netmap_sync_map_cpu(na, tag, map, sz, t) 1873 #define netmap_sync_map_dev(na, tag, map, sz, t) 1874 #endif /* NETMAP_LINUX_HAVE_DMASYNC */ 1875 1876 #endif /* linux */ 1877 1878 1879 /* 1880 * functions to map NIC to KRING indexes (n2k) and vice versa (k2n) 1881 */ 1882 static inline int 1883 netmap_idx_n2k(struct netmap_kring *kr, int idx) 1884 { 1885 int n = kr->nkr_num_slots; 1886 1887 if (likely(kr->nkr_hwofs == 0)) { 1888 return idx; 1889 } 1890 1891 idx += kr->nkr_hwofs; 1892 if (idx < 0) 1893 return idx + n; 1894 else if (idx < n) 1895 return idx; 1896 else 1897 return idx - n; 1898 } 1899 1900 1901 static inline int 1902 netmap_idx_k2n(struct netmap_kring *kr, int idx) 1903 { 1904 int n = kr->nkr_num_slots; 1905 1906 if (likely(kr->nkr_hwofs == 0)) { 1907 return idx; 1908 } 1909 1910 idx -= kr->nkr_hwofs; 1911 if (idx < 0) 1912 return idx + n; 1913 else if (idx < n) 1914 return idx; 1915 else 1916 return idx - n; 1917 } 1918 1919 1920 /* Entries of the look-up table. */ 1921 #ifdef __FreeBSD__ 1922 struct lut_entry { 1923 void *vaddr; /* virtual address. */ 1924 vm_paddr_t paddr; /* physical address. */ 1925 }; 1926 #else /* linux & _WIN32 */ 1927 /* dma-mapping in linux can assign a buffer a different address 1928 * depending on the device, so we need to have a separate 1929 * physical-address look-up table for each na. 1930 * We can still share the vaddrs, though, therefore we split 1931 * the lut_entry structure. 1932 */ 1933 struct lut_entry { 1934 void *vaddr; /* virtual address. */ 1935 }; 1936 1937 struct plut_entry { 1938 vm_paddr_t paddr; /* physical address. */ 1939 }; 1940 #endif /* linux & _WIN32 */ 1941 1942 struct netmap_obj_pool; 1943 1944 /* alignment for netmap buffers */ 1945 #define NM_BUF_ALIGN 64 1946 1947 /* 1948 * NMB return the virtual address of a buffer (buffer 0 on bad index) 1949 * PNMB also fills the physical address 1950 */ 1951 static inline void * 1952 NMB(struct netmap_adapter *na, struct netmap_slot *slot) 1953 { 1954 struct lut_entry *lut = na->na_lut.lut; 1955 uint32_t i = slot->buf_idx; 1956 return (unlikely(i >= na->na_lut.objtotal)) ? 1957 lut[0].vaddr : lut[i].vaddr; 1958 } 1959 1960 static inline void * 1961 PNMB(struct netmap_adapter *na, struct netmap_slot *slot, uint64_t *pp) 1962 { 1963 uint32_t i = slot->buf_idx; 1964 struct lut_entry *lut = na->na_lut.lut; 1965 struct plut_entry *plut = na->na_lut.plut; 1966 void *ret = (i >= na->na_lut.objtotal) ? lut[0].vaddr : lut[i].vaddr; 1967 1968 #ifdef _WIN32 1969 *pp = (i >= na->na_lut.objtotal) ? (uint64_t)plut[0].paddr.QuadPart : (uint64_t)plut[i].paddr.QuadPart; 1970 #else 1971 *pp = (i >= na->na_lut.objtotal) ? plut[0].paddr : plut[i].paddr; 1972 #endif 1973 return ret; 1974 } 1975 1976 static inline void 1977 nm_write_offset(struct netmap_kring *kring, 1978 struct netmap_slot *slot, uint64_t offset) 1979 { 1980 slot->ptr = (slot->ptr & ~kring->offset_mask) | 1981 (offset & kring->offset_mask); 1982 } 1983 1984 static inline uint64_t 1985 nm_get_offset(struct netmap_kring *kring, struct netmap_slot *slot) 1986 { 1987 uint64_t offset = (slot->ptr & kring->offset_mask); 1988 if (unlikely(offset > kring->offset_max)) 1989 offset = kring->offset_max; 1990 return offset; 1991 } 1992 1993 static inline void * 1994 NMB_O(struct netmap_kring *kring, struct netmap_slot *slot) 1995 { 1996 void *addr = NMB(kring->na, slot); 1997 return (char *)addr + nm_get_offset(kring, slot); 1998 } 1999 2000 static inline void * 2001 PNMB_O(struct netmap_kring *kring, struct netmap_slot *slot, uint64_t *pp) 2002 { 2003 void *addr = PNMB(kring->na, slot, pp); 2004 uint64_t offset = nm_get_offset(kring, slot); 2005 addr = (char *)addr + offset; 2006 *pp += offset; 2007 return addr; 2008 } 2009 2010 2011 /* 2012 * Structure associated to each netmap file descriptor. 2013 * It is created on open and left unbound (np_nifp == NULL). 2014 * A successful NIOCREGIF will set np_nifp and the first few fields; 2015 * this is protected by a global lock (NMG_LOCK) due to low contention. 2016 * 2017 * np_refs counts the number of references to the structure: one for the fd, 2018 * plus (on FreeBSD) one for each active mmap which we track ourselves 2019 * (linux automatically tracks them, but FreeBSD does not). 2020 * np_refs is protected by NMG_LOCK. 2021 * 2022 * Read access to the structure is lock free, because ni_nifp once set 2023 * can only go to 0 when nobody is using the entry anymore. Readers 2024 * must check that np_nifp != NULL before using the other fields. 2025 */ 2026 struct netmap_priv_d { 2027 struct netmap_if * volatile np_nifp; /* netmap if descriptor. */ 2028 2029 struct netmap_adapter *np_na; 2030 struct ifnet *np_ifp; 2031 uint32_t np_flags; /* from the ioctl */ 2032 u_int np_qfirst[NR_TXRX], 2033 np_qlast[NR_TXRX]; /* range of tx/rx rings to scan */ 2034 uint16_t np_txpoll; 2035 uint16_t np_kloop_state; /* use with NMG_LOCK held */ 2036 #define NM_SYNC_KLOOP_RUNNING (1 << 0) 2037 #define NM_SYNC_KLOOP_STOPPING (1 << 1) 2038 int np_sync_flags; /* to be passed to nm_sync */ 2039 2040 int np_refs; /* use with NMG_LOCK held */ 2041 2042 /* pointers to the selinfo to be used for selrecord. 2043 * Either the local or the global one depending on the 2044 * number of rings. 2045 */ 2046 NM_SELINFO_T *np_si[NR_TXRX]; 2047 2048 /* In the optional CSB mode, the user must specify the start address 2049 * of two arrays of Communication Status Block (CSB) entries, for the 2050 * two directions (kernel read application write, and kernel write 2051 * application read). 2052 * The number of entries must agree with the number of rings bound to 2053 * the netmap file descriptor. The entries corresponding to the TX 2054 * rings are laid out before the ones corresponding to the RX rings. 2055 * 2056 * Array of CSB entries for application --> kernel communication 2057 * (N entries). */ 2058 struct nm_csb_atok *np_csb_atok_base; 2059 /* Array of CSB entries for kernel --> application communication 2060 * (N entries). */ 2061 struct nm_csb_ktoa *np_csb_ktoa_base; 2062 2063 #ifdef linux 2064 struct file *np_filp; /* used by sync kloop */ 2065 #endif /* linux */ 2066 }; 2067 2068 struct netmap_priv_d *netmap_priv_new(void); 2069 void netmap_priv_delete(struct netmap_priv_d *); 2070 2071 static inline int nm_kring_pending(struct netmap_priv_d *np) 2072 { 2073 struct netmap_adapter *na = np->np_na; 2074 enum txrx t; 2075 int i; 2076 2077 for_rx_tx(t) { 2078 for (i = np->np_qfirst[t]; i < np->np_qlast[t]; i++) { 2079 struct netmap_kring *kring = NMR(na, t)[i]; 2080 if (kring->nr_mode != kring->nr_pending_mode) { 2081 return 1; 2082 } 2083 } 2084 } 2085 return 0; 2086 } 2087 2088 /* call with NMG_LOCK held */ 2089 static __inline int 2090 nm_si_user(struct netmap_priv_d *priv, enum txrx t) 2091 { 2092 return (priv->np_na != NULL && 2093 (priv->np_qlast[t] - priv->np_qfirst[t] > 1)); 2094 } 2095 2096 #ifdef WITH_PIPES 2097 int netmap_pipe_txsync(struct netmap_kring *txkring, int flags); 2098 int netmap_pipe_rxsync(struct netmap_kring *rxkring, int flags); 2099 int netmap_pipe_krings_create_both(struct netmap_adapter *na, 2100 struct netmap_adapter *ona); 2101 void netmap_pipe_krings_delete_both(struct netmap_adapter *na, 2102 struct netmap_adapter *ona); 2103 int netmap_pipe_reg_both(struct netmap_adapter *na, 2104 struct netmap_adapter *ona); 2105 #endif /* WITH_PIPES */ 2106 2107 #ifdef WITH_MONITOR 2108 2109 struct netmap_monitor_adapter { 2110 struct netmap_adapter up; 2111 2112 struct netmap_priv_d priv; 2113 uint32_t flags; 2114 }; 2115 2116 #endif /* WITH_MONITOR */ 2117 2118 2119 #ifdef WITH_GENERIC 2120 /* 2121 * generic netmap emulation for devices that do not have 2122 * native netmap support. 2123 */ 2124 int generic_netmap_attach(struct ifnet *ifp); 2125 int generic_rx_handler(struct ifnet *ifp, struct mbuf *m); 2126 2127 int nm_os_catch_rx(struct netmap_generic_adapter *gna, int intercept); 2128 int nm_os_catch_tx(struct netmap_generic_adapter *gna, int intercept); 2129 2130 int na_is_generic(struct netmap_adapter *na); 2131 2132 /* 2133 * the generic transmit routine is passed a structure to optionally 2134 * build a queue of descriptors, in an OS-specific way. 2135 * The payload is at addr, if non-null, and the routine should send or queue 2136 * the packet, returning 0 if successful, 1 on failure. 2137 * 2138 * At the end, if head is non-null, there will be an additional call 2139 * to the function with addr = NULL; this should tell the OS-specific 2140 * routine to send the queue and free any resources. Failure is ignored. 2141 */ 2142 struct nm_os_gen_arg { 2143 struct ifnet *ifp; 2144 void *m; /* os-specific mbuf-like object */ 2145 void *head, *tail; /* tailq, if the OS-specific routine needs to build one */ 2146 void *addr; /* payload of current packet */ 2147 u_int len; /* packet length */ 2148 u_int ring_nr; /* packet length */ 2149 u_int qevent; /* in txqdisc mode, place an event on this mbuf */ 2150 }; 2151 2152 int nm_os_generic_xmit_frame(struct nm_os_gen_arg *); 2153 int nm_os_generic_find_num_desc(struct ifnet *ifp, u_int *tx, u_int *rx); 2154 void nm_os_generic_find_num_queues(struct ifnet *ifp, u_int *txq, u_int *rxq); 2155 void nm_os_generic_set_features(struct netmap_generic_adapter *gna); 2156 2157 static inline struct ifnet* 2158 netmap_generic_getifp(struct netmap_generic_adapter *gna) 2159 { 2160 if (gna->prev) 2161 return gna->prev->ifp; 2162 2163 return gna->up.up.ifp; 2164 } 2165 2166 void netmap_generic_irq(struct netmap_adapter *na, u_int q, u_int *work_done); 2167 2168 //#define RATE_GENERIC /* Enables communication statistics for generic. */ 2169 #ifdef RATE_GENERIC 2170 void generic_rate(int txp, int txs, int txi, int rxp, int rxs, int rxi); 2171 #else 2172 #define generic_rate(txp, txs, txi, rxp, rxs, rxi) 2173 #endif 2174 2175 /* 2176 * netmap_mitigation API. This is used by the generic adapter 2177 * to reduce the number of interrupt requests/selwakeup 2178 * to clients on incoming packets. 2179 */ 2180 void nm_os_mitigation_init(struct nm_generic_mit *mit, int idx, 2181 struct netmap_adapter *na); 2182 void nm_os_mitigation_start(struct nm_generic_mit *mit); 2183 void nm_os_mitigation_restart(struct nm_generic_mit *mit); 2184 int nm_os_mitigation_active(struct nm_generic_mit *mit); 2185 void nm_os_mitigation_cleanup(struct nm_generic_mit *mit); 2186 #else /* !WITH_GENERIC */ 2187 #define generic_netmap_attach(ifp) (EOPNOTSUPP) 2188 #define na_is_generic(na) (0) 2189 #endif /* WITH_GENERIC */ 2190 2191 /* Shared declarations for the VALE switch. */ 2192 2193 /* 2194 * Each transmit queue accumulates a batch of packets into 2195 * a structure before forwarding. Packets to the same 2196 * destination are put in a list using ft_next as a link field. 2197 * ft_frags and ft_next are valid only on the first fragment. 2198 */ 2199 struct nm_bdg_fwd { /* forwarding entry for a bridge */ 2200 void *ft_buf; /* netmap or indirect buffer */ 2201 uint8_t ft_frags; /* how many fragments (only on 1st frag) */ 2202 uint16_t ft_offset; /* dst port (unused) */ 2203 uint16_t ft_flags; /* flags, e.g. indirect */ 2204 uint16_t ft_len; /* src fragment len */ 2205 uint16_t ft_next; /* next packet to same destination */ 2206 }; 2207 2208 /* struct 'virtio_net_hdr' from linux. */ 2209 struct nm_vnet_hdr { 2210 #define VIRTIO_NET_HDR_F_NEEDS_CSUM 1 /* Use csum_start, csum_offset */ 2211 #define VIRTIO_NET_HDR_F_DATA_VALID 2 /* Csum is valid */ 2212 uint8_t flags; 2213 #define VIRTIO_NET_HDR_GSO_NONE 0 /* Not a GSO frame */ 2214 #define VIRTIO_NET_HDR_GSO_TCPV4 1 /* GSO frame, IPv4 TCP (TSO) */ 2215 #define VIRTIO_NET_HDR_GSO_UDP 3 /* GSO frame, IPv4 UDP (UFO) */ 2216 #define VIRTIO_NET_HDR_GSO_TCPV6 4 /* GSO frame, IPv6 TCP */ 2217 #define VIRTIO_NET_HDR_GSO_ECN 0x80 /* TCP has ECN set */ 2218 uint8_t gso_type; 2219 uint16_t hdr_len; 2220 uint16_t gso_size; 2221 uint16_t csum_start; 2222 uint16_t csum_offset; 2223 }; 2224 2225 #define WORST_CASE_GSO_HEADER (14+40+60) /* IPv6 + TCP */ 2226 2227 /* Private definitions for IPv4, IPv6, UDP and TCP headers. */ 2228 2229 struct nm_iphdr { 2230 uint8_t version_ihl; 2231 uint8_t tos; 2232 uint16_t tot_len; 2233 uint16_t id; 2234 uint16_t frag_off; 2235 uint8_t ttl; 2236 uint8_t protocol; 2237 uint16_t check; 2238 uint32_t saddr; 2239 uint32_t daddr; 2240 /*The options start here. */ 2241 }; 2242 2243 struct nm_tcphdr { 2244 uint16_t source; 2245 uint16_t dest; 2246 uint32_t seq; 2247 uint32_t ack_seq; 2248 uint8_t doff; /* Data offset + Reserved */ 2249 uint8_t flags; 2250 uint16_t window; 2251 uint16_t check; 2252 uint16_t urg_ptr; 2253 }; 2254 2255 struct nm_udphdr { 2256 uint16_t source; 2257 uint16_t dest; 2258 uint16_t len; 2259 uint16_t check; 2260 }; 2261 2262 struct nm_ipv6hdr { 2263 uint8_t priority_version; 2264 uint8_t flow_lbl[3]; 2265 2266 uint16_t payload_len; 2267 uint8_t nexthdr; 2268 uint8_t hop_limit; 2269 2270 uint8_t saddr[16]; 2271 uint8_t daddr[16]; 2272 }; 2273 2274 /* Type used to store a checksum (in host byte order) that hasn't been 2275 * folded yet. 2276 */ 2277 #define rawsum_t uint32_t 2278 2279 rawsum_t nm_os_csum_raw(uint8_t *data, size_t len, rawsum_t cur_sum); 2280 uint16_t nm_os_csum_ipv4(struct nm_iphdr *iph); 2281 void nm_os_csum_tcpudp_ipv4(struct nm_iphdr *iph, void *data, 2282 size_t datalen, uint16_t *check); 2283 void nm_os_csum_tcpudp_ipv6(struct nm_ipv6hdr *ip6h, void *data, 2284 size_t datalen, uint16_t *check); 2285 uint16_t nm_os_csum_fold(rawsum_t cur_sum); 2286 2287 void bdg_mismatch_datapath(struct netmap_vp_adapter *na, 2288 struct netmap_vp_adapter *dst_na, 2289 const struct nm_bdg_fwd *ft_p, 2290 struct netmap_ring *dst_ring, 2291 u_int *j, u_int lim, u_int *howmany); 2292 2293 /* persistent virtual port routines */ 2294 int nm_os_vi_persist(const char *, struct ifnet **); 2295 void nm_os_vi_detach(struct ifnet *); 2296 void nm_os_vi_init_index(void); 2297 2298 /* 2299 * kernel thread routines 2300 */ 2301 struct nm_kctx; /* OS-specific kernel context - opaque */ 2302 typedef void (*nm_kctx_worker_fn_t)(void *data); 2303 2304 /* kthread configuration */ 2305 struct nm_kctx_cfg { 2306 long type; /* kthread type/identifier */ 2307 nm_kctx_worker_fn_t worker_fn; /* worker function */ 2308 void *worker_private;/* worker parameter */ 2309 int attach_user; /* attach kthread to user process */ 2310 }; 2311 /* kthread configuration */ 2312 struct nm_kctx *nm_os_kctx_create(struct nm_kctx_cfg *cfg, 2313 void *opaque); 2314 int nm_os_kctx_worker_start(struct nm_kctx *); 2315 void nm_os_kctx_worker_stop(struct nm_kctx *); 2316 void nm_os_kctx_destroy(struct nm_kctx *); 2317 void nm_os_kctx_worker_setaff(struct nm_kctx *, int); 2318 u_int nm_os_ncpus(void); 2319 2320 int netmap_sync_kloop(struct netmap_priv_d *priv, 2321 struct nmreq_header *hdr); 2322 int netmap_sync_kloop_stop(struct netmap_priv_d *priv); 2323 2324 #ifdef WITH_PTNETMAP 2325 /* ptnetmap guest routines */ 2326 2327 /* 2328 * ptnetmap_memdev routines used to talk with ptnetmap_memdev device driver 2329 */ 2330 struct ptnetmap_memdev; 2331 int nm_os_pt_memdev_iomap(struct ptnetmap_memdev *, vm_paddr_t *, void **, 2332 uint64_t *); 2333 void nm_os_pt_memdev_iounmap(struct ptnetmap_memdev *); 2334 uint32_t nm_os_pt_memdev_ioread(struct ptnetmap_memdev *, unsigned int); 2335 2336 /* 2337 * netmap adapter for guest ptnetmap ports 2338 */ 2339 struct netmap_pt_guest_adapter { 2340 /* The netmap adapter to be used by netmap applications. 2341 * This field must be the first, to allow upcast. */ 2342 struct netmap_hw_adapter hwup; 2343 2344 /* The netmap adapter to be used by the driver. */ 2345 struct netmap_hw_adapter dr; 2346 2347 /* Reference counter to track users of backend netmap port: the 2348 * network stack and netmap clients. 2349 * Used to decide when we need (de)allocate krings/rings and 2350 * start (stop) ptnetmap kthreads. */ 2351 int backend_users; 2352 2353 }; 2354 2355 int netmap_pt_guest_attach(struct netmap_adapter *na, 2356 unsigned int nifp_offset, 2357 unsigned int memid); 2358 bool netmap_pt_guest_txsync(struct nm_csb_atok *atok, 2359 struct nm_csb_ktoa *ktoa, 2360 struct netmap_kring *kring, int flags); 2361 bool netmap_pt_guest_rxsync(struct nm_csb_atok *atok, 2362 struct nm_csb_ktoa *ktoa, 2363 struct netmap_kring *kring, int flags); 2364 int ptnet_nm_krings_create(struct netmap_adapter *na); 2365 void ptnet_nm_krings_delete(struct netmap_adapter *na); 2366 void ptnet_nm_dtor(struct netmap_adapter *na); 2367 2368 /* Helper function wrapping nm_sync_kloop_appl_read(). */ 2369 static inline void 2370 ptnet_sync_tail(struct nm_csb_ktoa *ktoa, struct netmap_kring *kring) 2371 { 2372 struct netmap_ring *ring = kring->ring; 2373 2374 /* Update hwcur and hwtail as known by the host. */ 2375 nm_sync_kloop_appl_read(ktoa, &kring->nr_hwtail, &kring->nr_hwcur); 2376 2377 /* nm_sync_finalize */ 2378 ring->tail = kring->rtail = kring->nr_hwtail; 2379 } 2380 #endif /* WITH_PTNETMAP */ 2381 2382 #ifdef __FreeBSD__ 2383 /* 2384 * FreeBSD mbuf allocator/deallocator in emulation mode: 2385 * 2386 * We allocate mbufs with m_gethdr(), since the mbuf header is needed 2387 * by the driver. We also attach a customly-provided external storage, 2388 * which in this case is a netmap buffer. When calling m_extadd(), however 2389 * we pass a NULL address, since the real address (and length) will be 2390 * filled in by nm_os_generic_xmit_frame() right before calling 2391 * if_transmit(). 2392 * 2393 * The dtor function does nothing, however we need it since mb_free_ext() 2394 * has a KASSERT(), checking that the mbuf dtor function is not NULL. 2395 */ 2396 2397 static void void_mbuf_dtor(struct mbuf *m) { } 2398 2399 #define SET_MBUF_DESTRUCTOR(m, fn) do { \ 2400 (m)->m_ext.ext_free = (fn != NULL) ? \ 2401 (void *)fn : (void *)void_mbuf_dtor; \ 2402 } while (0) 2403 2404 static inline struct mbuf * 2405 nm_os_get_mbuf(struct ifnet *ifp, int len) 2406 { 2407 struct mbuf *m; 2408 2409 (void)ifp; 2410 (void)len; 2411 2412 m = m_gethdr(M_NOWAIT, MT_DATA); 2413 if (m == NULL) { 2414 return m; 2415 } 2416 2417 m_extadd(m, NULL /* buf */, 0 /* size */, void_mbuf_dtor, 2418 NULL, NULL, 0, EXT_NET_DRV); 2419 2420 return m; 2421 } 2422 2423 #endif /* __FreeBSD__ */ 2424 2425 struct nmreq_option * nmreq_getoption(struct nmreq_header *, uint16_t); 2426 2427 int netmap_init_bridges(void); 2428 void netmap_uninit_bridges(void); 2429 2430 /* Functions to read and write CSB fields from the kernel. */ 2431 #if defined (linux) 2432 #define CSB_READ(csb, field, r) (get_user(r, &csb->field)) 2433 #define CSB_WRITE(csb, field, v) (put_user(v, &csb->field)) 2434 #else /* ! linux */ 2435 #define CSB_READ(csb, field, r) (r = fuword32(&csb->field)) 2436 #define CSB_WRITE(csb, field, v) (suword32(&csb->field, v)) 2437 #endif /* ! linux */ 2438 2439 /* some macros that may not be defined */ 2440 #ifndef ETH_HLEN 2441 #define ETH_HLEN 6 2442 #endif 2443 #ifndef ETH_FCS_LEN 2444 #define ETH_FCS_LEN 4 2445 #endif 2446 #ifndef VLAN_HLEN 2447 #define VLAN_HLEN 4 2448 #endif 2449 2450 #endif /* _NET_NETMAP_KERN_H_ */ 2451