2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo
5 * Copyright (C) 2013-2016 Universita` di Pisa
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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
33 * The header contains the definitions of constants and function
34 * prototypes used only in kernelspace.
37 #ifndef _NET_NETMAP_KERN_H_
38 #define _NET_NETMAP_KERN_H_
42 #if defined(CONFIG_NETMAP_EXTMEM)
45 #if defined(CONFIG_NETMAP_VALE)
48 #if defined(CONFIG_NETMAP_PIPE)
51 #if defined(CONFIG_NETMAP_MONITOR)
54 #if defined(CONFIG_NETMAP_GENERIC)
57 #if defined(CONFIG_NETMAP_PTNETMAP)
60 #if defined(CONFIG_NETMAP_SINK)
63 #if defined(CONFIG_NETMAP_NULL)
67 #elif defined (_WIN32)
68 #define WITH_VALE // comment out to disable VALE support
74 #else /* neither linux nor windows */
75 #define WITH_VALE // comment out to disable VALE support
83 #if defined(__FreeBSD__)
84 #include <sys/selinfo.h>
86 #define likely(x) __builtin_expect((long)!!(x), 1L)
87 #define unlikely(x) __builtin_expect((long)!!(x), 0L)
90 #define NM_LOCK_T struct mtx /* low level spinlock, used to protect queues */
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)
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)
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)
113 #if __FreeBSD_version >= 1100030
114 #define WNA(_ifp) (_ifp)->if_netmap
115 #else /* older FreeBSD */
116 #define WNA(_ifp) (_ifp)->if_pspare[0]
117 #endif /* older FreeBSD */
119 #if __FreeBSD_version >= 1100005
120 struct netmap_adapter *netmap_getna(if_t ifp);
123 #if __FreeBSD_version >= 1100027
124 #define MBUF_REFCNT(m) ((m)->m_ext.ext_count)
125 #define SET_MBUF_REFCNT(m, x) (m)->m_ext.ext_count = x
127 #define MBUF_REFCNT(m) ((m)->m_ext.ref_cnt ? *((m)->m_ext.ref_cnt) : -1)
128 #define SET_MBUF_REFCNT(m, x) *((m)->m_ext.ref_cnt) = x
131 #define MBUF_QUEUED(m) 1
134 /* Support for select(2) and poll(2). */
136 /* Support for kqueue(9). See comments in netmap_freebsd.c */
137 struct taskqueue *ntfytq;
138 struct task ntfytask;
146 /* Not used in FreeBSD. */
149 #define NM_BNS_GET(b)
150 #define NM_BNS_PUT(b)
152 #elif defined (linux)
154 #define NM_LOCK_T safe_spinlock_t // see bsd_glue.h
155 #define NM_SELINFO_T wait_queue_head_t
156 #define MBUF_LEN(m) ((m)->len)
157 #define MBUF_TRANSMIT(na, ifp, m) \
159 /* Avoid infinite recursion with generic. */ \
160 m->priority = NM_MAGIC_PRIORITY_TX; \
161 (((struct net_device_ops *)(na)->if_transmit)->ndo_start_xmit(m, ifp)); \
165 /* See explanation in nm_os_generic_xmit_frame. */
166 #define GEN_TX_MBUF_IFP(m) ((struct ifnet *)skb_shinfo(m)->destructor_arg)
168 #define NM_ATOMIC_T volatile long unsigned int
170 #define NM_MTX_T struct mutex /* OS-specific sleepable lock */
171 #define NM_MTX_INIT(m) mutex_init(&(m))
172 #define NM_MTX_DESTROY(m) do { (void)(m); } while (0)
173 #define NM_MTX_LOCK(m) mutex_lock(&(m))
174 #define NM_MTX_UNLOCK(m) mutex_unlock(&(m))
175 #define NM_MTX_ASSERT(m) mutex_is_locked(&(m))
179 #endif /* DEV_NETMAP */
181 #elif defined (__APPLE__)
183 #warning apple support is incomplete.
184 #define likely(x) __builtin_expect(!!(x), 1)
185 #define unlikely(x) __builtin_expect(!!(x), 0)
186 #define NM_LOCK_T IOLock *
187 #define NM_SELINFO_T struct selinfo
188 #define MBUF_LEN(m) ((m)->m_pkthdr.len)
190 #elif defined (_WIN32)
191 #include "../../../WINDOWS/win_glue.h"
193 #define NM_SELRECORD_T IO_STACK_LOCATION
194 #define NM_SELINFO_T win_SELINFO // see win_glue.h
195 #define NM_LOCK_T win_spinlock_t // see win_glue.h
196 #define NM_MTX_T KGUARDED_MUTEX /* OS-specific mutex (sleepable) */
198 #define NM_MTX_INIT(m) KeInitializeGuardedMutex(&m);
199 #define NM_MTX_DESTROY(m) do { (void)(m); } while (0)
200 #define NM_MTX_LOCK(m) KeAcquireGuardedMutex(&(m))
201 #define NM_MTX_UNLOCK(m) KeReleaseGuardedMutex(&(m))
202 #define NM_MTX_ASSERT(m) assert(&m.Count>0)
204 //These linknames are for the NDIS driver
205 #define NETMAP_NDIS_LINKNAME_STRING L"\\DosDevices\\NMAPNDIS"
206 #define NETMAP_NDIS_NTDEVICE_STRING L"\\Device\\NMAPNDIS"
208 //Definition of internal driver-to-driver ioctl codes
209 #define NETMAP_KERNEL_XCHANGE_POINTERS _IO('i', 180)
210 #define NETMAP_KERNEL_SEND_SHUTDOWN_SIGNAL _IO_direct('i', 195)
212 typedef struct hrtimer{
218 /* MSVC does not have likely/unlikely support */
220 #define likely(x) (x)
221 #define unlikely(x) (x)
223 #define likely(x) __builtin_expect((long)!!(x), 1L)
224 #define unlikely(x) __builtin_expect((long)!!(x), 0L)
229 #error unsupported platform
231 #endif /* end - platform-specific code */
233 #ifndef _WIN32 /* support for emulated sysctl */
238 #define NM_ACCESS_ONCE(x) (*(volatile __typeof__(x) *)&(x))
240 #define NMG_LOCK_T NM_MTX_T
241 #define NMG_LOCK_INIT() NM_MTX_INIT(netmap_global_lock)
242 #define NMG_LOCK_DESTROY() NM_MTX_DESTROY(netmap_global_lock)
243 #define NMG_LOCK() NM_MTX_LOCK(netmap_global_lock)
244 #define NMG_UNLOCK() NM_MTX_UNLOCK(netmap_global_lock)
245 #define NMG_LOCK_ASSERT() NM_MTX_ASSERT(netmap_global_lock)
247 #if defined(__FreeBSD__)
248 #define nm_prerr_int printf
249 #define nm_prinf_int printf
250 #elif defined (_WIN32)
251 #define nm_prerr_int DbgPrint
252 #define nm_prinf_int DbgPrint
254 #define nm_prerr_int(fmt, arg...) printk(KERN_ERR fmt, ##arg)
255 #define nm_prinf_int(fmt, arg...) printk(KERN_INFO fmt, ##arg)
258 #define nm_prinf(format, ...) \
260 struct timeval __xxts; \
261 microtime(&__xxts); \
262 nm_prinf_int("%03d.%06d [%4d] %-25s " format "\n",\
263 (int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec, \
264 __LINE__, __FUNCTION__, ##__VA_ARGS__); \
267 #define nm_prerr(format, ...) \
269 struct timeval __xxts; \
270 microtime(&__xxts); \
271 nm_prerr_int("%03d.%06d [%4d] %-25s " format "\n",\
272 (int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec, \
273 __LINE__, __FUNCTION__, ##__VA_ARGS__); \
276 /* Disabled printf (used to be nm_prdis). */
277 #define nm_prdis(format, ...)
279 /* Rate limited, lps indicates how many per second. */
280 #define nm_prlim(lps, format, ...) \
282 static int t0, __cnt; \
283 if (t0 != time_second) { \
288 nm_prinf(format, ##__VA_ARGS__); \
291 struct netmap_adapter;
294 struct netmap_priv_d;
297 /* os-specific NM_SELINFO_T initialization/destruction functions */
298 int nm_os_selinfo_init(NM_SELINFO_T *, const char *name);
299 void nm_os_selinfo_uninit(NM_SELINFO_T *);
301 const char *nm_dump_buf(char *p, int len, int lim, char *dst);
303 void nm_os_selwakeup(NM_SELINFO_T *si);
304 void nm_os_selrecord(NM_SELRECORD_T *sr, NM_SELINFO_T *si);
306 int nm_os_ifnet_init(void);
307 void nm_os_ifnet_fini(void);
308 void nm_os_ifnet_lock(void);
309 void nm_os_ifnet_unlock(void);
311 unsigned nm_os_ifnet_mtu(struct ifnet *ifp);
313 void nm_os_get_module(void);
314 void nm_os_put_module(void);
316 void netmap_make_zombie(struct ifnet *);
317 void netmap_undo_zombie(struct ifnet *);
319 /* os independent alloc/realloc/free */
320 void *nm_os_malloc(size_t);
321 void *nm_os_vmalloc(size_t);
322 void *nm_os_realloc(void *, size_t new_size, size_t old_size);
323 void nm_os_free(void *);
324 void nm_os_vfree(void *);
326 /* os specific attach/detach enter/exit-netmap-mode routines */
327 void nm_os_onattach(struct ifnet *);
328 void nm_os_ondetach(struct ifnet *);
329 void nm_os_onenter(struct ifnet *);
330 void nm_os_onexit(struct ifnet *);
332 /* passes a packet up to the host stack.
333 * If the packet is sent (or dropped) immediately it returns NULL,
334 * otherwise it links the packet to prev and returns m.
335 * In this case, a final call with m=NULL and prev != NULL will send up
336 * the entire chain to the host stack.
338 void *nm_os_send_up(struct ifnet *, struct mbuf *m, struct mbuf *prev);
340 int nm_os_mbuf_has_seg_offld(struct mbuf *m);
341 int nm_os_mbuf_has_csum_offld(struct mbuf *m);
343 #include "netmap_mbq.h"
345 extern NMG_LOCK_T netmap_global_lock;
347 enum txrx { NR_RX = 0, NR_TX = 1, NR_TXRX };
349 static __inline const char*
350 nm_txrx2str(enum txrx t)
352 return (t== NR_RX ? "RX" : "TX");
355 static __inline enum txrx
356 nm_txrx_swap(enum txrx t)
358 return (t== NR_RX ? NR_TX : NR_RX);
361 #define for_rx_tx(t) for ((t) = 0; (t) < NR_TXRX; (t)++)
364 struct netmap_zmon_list {
365 struct netmap_kring *next;
366 struct netmap_kring *prev;
368 #endif /* WITH_MONITOR */
371 * private, kernel view of a ring. Keeps track of the status of
372 * a ring across system calls.
374 * nr_hwcur index of the next buffer to refill.
375 * It corresponds to ring->head
376 * at the time the system call returns.
378 * nr_hwtail index of the first buffer owned by the kernel.
379 * On RX, hwcur->hwtail are receive buffers
380 * not yet released. hwcur is advanced following
381 * ring->head, hwtail is advanced on incoming packets,
382 * and a wakeup is generated when hwtail passes ring->cur
383 * On TX, hwcur->rcur have been filled by the sender
384 * but not sent yet to the NIC; rcur->hwtail are available
385 * for new transmissions, and hwtail->hwcur-1 are pending
386 * transmissions not yet acknowledged.
388 * The indexes in the NIC and netmap rings are offset by nkr_hwofs slots.
389 * This is so that, on a reset, buffers owned by userspace are not
390 * modified by the kernel. In particular:
391 * RX rings: the next empty buffer (hwtail + hwofs) coincides with
392 * the next empty buffer as known by the hardware (next_to_check or so).
393 * TX rings: hwcur + hwofs coincides with next_to_send
395 * The following fields are used to implement lock-free copy of packets
396 * from input to output ports in VALE switch:
397 * nkr_hwlease buffer after the last one being copied.
398 * A writer in nm_bdg_flush reserves N buffers
399 * from nr_hwlease, advances it, then does the
400 * copy outside the lock.
401 * In RX rings (used for VALE ports),
402 * nkr_hwtail <= nkr_hwlease < nkr_hwcur+N-1
403 * In TX rings (used for NIC or host stack ports)
404 * nkr_hwcur <= nkr_hwlease < nkr_hwtail
405 * nkr_leases array of nkr_num_slots where writers can report
406 * completion of their block. NR_NOSLOT (~0) indicates
407 * that the writer has not finished yet
408 * nkr_lease_idx index of next free slot in nr_leases, to be assigned
410 * The kring is manipulated by txsync/rxsync and generic netmap function.
412 * Concurrent rxsync or txsync on the same ring are prevented through
413 * by nm_kr_(try)lock() which in turn uses nr_busy. This is all we need
414 * for NIC rings, and for TX rings attached to the host stack.
416 * RX rings attached to the host stack use an mbq (rx_queue) on both
417 * rxsync_from_host() and netmap_transmit(). The mbq is protected
418 * by its internal lock.
420 * RX rings attached to the VALE switch are accessed by both senders
421 * and receiver. They are protected through the q_lock on the RX ring.
423 struct netmap_kring {
424 struct netmap_ring *ring;
426 uint32_t nr_hwcur; /* should be nr_hwhead */
430 * Copies of values in user rings, so we do not need to look
431 * at the ring (which could be modified). These are set in the
432 * *sync_prologue()/finalize() routines.
438 uint32_t nr_kflags; /* private driver flags */
439 #define NKR_PENDINTR 0x1 // Pending interrupt.
440 #define NKR_EXCLUSIVE 0x2 /* exclusive binding */
441 #define NKR_FORWARD 0x4 /* (host ring only) there are
444 #define NKR_NEEDRING 0x8 /* ring needed even if users==0
445 * (used internally by pipes and
446 * by ptnetmap host ports)
448 #define NKR_NOINTR 0x10 /* don't use interrupts on this ring */
449 #define NKR_FAKERING 0x20 /* don't allocate/free buffers */
452 uint32_t nr_pending_mode;
453 #define NKR_NETMAP_OFF 0x0
454 #define NKR_NETMAP_ON 0x1
456 uint32_t nkr_num_slots;
459 * On a NIC reset, the NIC ring indexes may be reset but the
460 * indexes in the netmap rings remain the same. nkr_hwofs
461 * keeps track of the offset between the two.
463 * Moreover, during reset, we can restore only the subset of
464 * the NIC ring that corresponds to the kernel-owned part of
465 * the netmap ring. The rest of the slots must be restored
466 * by the *sync routines when the user releases more slots.
467 * The nkr_to_refill field keeps track of the number of slots
468 * that still need to be restored.
471 int32_t nkr_to_refill;
473 /* last_reclaim is opaque marker to help reduce the frequency
474 * of operations such as reclaiming tx buffers. A possible use
475 * is set it to ticks and do the reclaim only once per tick.
477 uint64_t last_reclaim;
480 NM_SELINFO_T si; /* poll/select wait queue */
481 NM_LOCK_T q_lock; /* protects kring and ring. */
482 NM_ATOMIC_T nr_busy; /* prevent concurrent syscalls */
484 /* the adapter the owns this kring */
485 struct netmap_adapter *na;
487 /* the adapter that wants to be notified when this kring has
488 * new slots available. This is usually the same as the above,
489 * but wrappers may let it point to themselves
491 struct netmap_adapter *notify_na;
493 /* The following fields are for VALE switch support */
494 struct nm_bdg_fwd *nkr_ft;
495 uint32_t *nkr_leases;
496 #define NR_NOSLOT ((uint32_t)~0) /* used in nkr_*lease* */
497 uint32_t nkr_hwlease;
498 uint32_t nkr_lease_idx;
500 /* while nkr_stopped is set, no new [tr]xsync operations can
501 * be started on this kring.
502 * This is used by netmap_disable_all_rings()
503 * to find a synchronization point where critical data
504 * structures pointed to by the kring can be added or removed
506 volatile int nkr_stopped;
508 /* Support for adapters without native netmap support.
509 * On tx rings we preallocate an array of tx buffers
510 * (same size as the netmap ring), on rx rings we
511 * store incoming mbufs in a queue that is drained by
514 struct mbuf **tx_pool;
515 struct mbuf *tx_event; /* TX event used as a notification */
516 NM_LOCK_T tx_event_lock; /* protects the tx_event mbuf */
517 struct mbq rx_queue; /* intercepted rx mbufs. */
519 uint32_t users; /* existing bindings for this ring */
521 uint32_t ring_id; /* kring identifier */
522 enum txrx tx; /* kind of ring (tx or rx) */
523 char name[64]; /* diagnostic */
525 /* [tx]sync callback for this kring.
526 * The default nm_kring_create callback (netmap_krings_create)
527 * sets the nm_sync callback of each hardware tx(rx) kring to
528 * the corresponding nm_txsync(nm_rxsync) taken from the
529 * netmap_adapter; moreover, it sets the sync callback
530 * of the host tx(rx) ring to netmap_txsync_to_host
531 * (netmap_rxsync_from_host).
533 * Overrides: the above configuration is not changed by
534 * any of the nm_krings_create callbacks.
536 int (*nm_sync)(struct netmap_kring *kring, int flags);
537 int (*nm_notify)(struct netmap_kring *kring, int flags);
540 struct netmap_kring *pipe; /* if this is a pipe ring,
541 * pointer to the other end
543 uint32_t pipe_tail; /* hwtail updated by the other end */
544 #endif /* WITH_PIPES */
546 /* mask for the offset-related part of the ptr field in the slots */
547 uint64_t offset_mask;
548 /* maximum user-specified offset, as stipulated at bind time.
549 * Larger offset requests will be silently capped to offset_max.
552 /* minimum gap between two consecutive offsets into the same
553 * buffer, as stipulated at bind time. This is used to choose
554 * the hwbuf_len, but is not otherwise checked for compliance
559 /* size of hardware buffer. This may be less than the size of
560 * the netmap buffers because of non-zero offsets, or because
561 * the netmap buffer size exceeds the capability of the hardware.
565 /* required alignment (in bytes) for the buffers used by this ring.
566 * Netmap buffers are aligned to cachelines, which should suffice
567 * for most NICs. If the user is passing offsets, though, we need
568 * to check that the resulting buf address complies with any
569 * alignment restriction.
573 /* hardware specific logic for the selection of the hwbuf_len */
574 int (*nm_bufcfg)(struct netmap_kring *kring, uint64_t target);
576 int (*save_notify)(struct netmap_kring *kring, int flags);
579 /* array of krings that are monitoring this kring */
580 struct netmap_kring **monitors;
581 uint32_t max_monitors; /* current size of the monitors array */
582 uint32_t n_monitors; /* next unused entry in the monitor array */
583 uint32_t mon_pos[NR_TXRX]; /* index of this ring in the monitored ring array */
584 uint32_t mon_tail; /* last seen slot on rx */
586 /* circular list of zero-copy monitors */
587 struct netmap_zmon_list zmon_list[NR_TXRX];
590 * Monitors work by intercepting the sync and notify callbacks of the
591 * monitored krings. This is implemented by replacing the pointers
592 * above and saving the previous ones in mon_* pointers below
594 int (*mon_sync)(struct netmap_kring *kring, int flags);
595 int (*mon_notify)(struct netmap_kring *kring, int flags);
600 __declspec(align(64));
602 __attribute__((__aligned__(64)));
605 /* return 1 iff the kring needs to be turned on */
607 nm_kring_pending_on(struct netmap_kring *kring)
609 return kring->nr_pending_mode == NKR_NETMAP_ON &&
610 kring->nr_mode == NKR_NETMAP_OFF;
613 /* return 1 iff the kring needs to be turned off */
615 nm_kring_pending_off(struct netmap_kring *kring)
617 return kring->nr_pending_mode == NKR_NETMAP_OFF &&
618 kring->nr_mode == NKR_NETMAP_ON;
621 /* return the next index, with wraparound */
622 static inline uint32_t
623 nm_next(uint32_t i, uint32_t lim)
625 return unlikely (i == lim) ? 0 : i + 1;
629 /* return the previous index, with wraparound */
630 static inline uint32_t
631 nm_prev(uint32_t i, uint32_t lim)
633 return unlikely (i == 0) ? lim : i - 1;
639 * Here is the layout for the Rx and Tx rings.
643 +-----------------+ +-----------------+
646 +-----------------+ +-----------------+
647 head->| owned by user |<-hwcur | not sent to nic |<-hwcur
649 +-----------------+ | |
650 cur->| available to | | |
651 | user, not read | +-----------------+
652 | yet | cur->| (being |
655 +-----------------+ + ------ +
656 tail->| |<-hwtail | |<-hwlease
657 | (being | ... | | ...
658 | prepared) | ... | | ...
659 +-----------------+ ... | | ...
660 | |<-hwlease +-----------------+
661 | | tail->| |<-hwtail
665 +-----------------+ +-----------------+
667 * The cur/tail (user view) and hwcur/hwtail (kernel view)
668 * are used in the normal operation of the card.
670 * When a ring is the output of a switch port (Rx ring for
671 * a VALE port, Tx ring for the host stack or NIC), slots
672 * are reserved in blocks through 'hwlease' which points
673 * to the next unused slot.
674 * On an Rx ring, hwlease is always after hwtail,
675 * and completions cause hwtail to advance.
676 * On a Tx ring, hwlease is always between cur and hwtail,
677 * and completions cause cur to advance.
679 * nm_kr_space() returns the maximum number of slots that
681 * nm_kr_lease() reserves the required number of buffers,
682 * advances nkr_hwlease and also returns an entry in
683 * a circular array where completions should be reported.
688 #define plut_entry lut_entry
692 struct lut_entry *lut;
693 struct plut_entry *plut;
694 uint32_t objtotal; /* max buffer index */
695 uint32_t objsize; /* buffer size */
698 struct netmap_vp_adapter; // forward
701 /* Struct to be filled by nm_config callbacks. */
702 struct nm_config_info {
703 unsigned num_tx_rings;
704 unsigned num_rx_rings;
705 unsigned num_tx_descs;
706 unsigned num_rx_descs;
707 unsigned rx_buf_maxsize;
711 * default type for the magic field.
712 * May be overridden in glue code.
715 #define NM_OS_MAGIC uint32_t
716 #endif /* !NM_OS_MAGIC */
719 * The "struct netmap_adapter" extends the "struct adapter"
720 * (or equivalent) device descriptor.
721 * It contains all base fields needed to support netmap operation.
722 * There are in fact different types of netmap adapters
723 * (native, generic, VALE switch...) so a netmap_adapter is
724 * just the first field in the derived type.
726 struct netmap_adapter {
728 * On linux we do not have a good way to tell if an interface
729 * is netmap-capable. So we always use the following trick:
730 * NA(ifp) points here, and the first entry (which hopefully
731 * always exists and is at least 32 bits) contains a magic
732 * value which we can use to detect that the interface is good.
735 uint32_t na_flags; /* enabled, and other flags */
736 #define NAF_SKIP_INTR 1 /* use the regular interrupt handler.
737 * useful during initialization
739 #define NAF_SW_ONLY 2 /* forward packets only to sw adapter */
740 #define NAF_BDG_MAYSLEEP 4 /* the bridge is allowed to sleep when
741 * forwarding packets coming from this
744 #define NAF_MEM_OWNER 8 /* the adapter uses its own memory area
745 * that cannot be changed
747 #define NAF_NATIVE 16 /* the adapter is native.
748 * Virtual ports (non persistent vale ports,
749 * pipes, monitors...) should never use
752 #define NAF_NETMAP_ON 32 /* netmap is active (either native or
753 * emulated). Where possible (e.g. FreeBSD)
754 * IFCAP_NETMAP also mirrors this flag.
756 #define NAF_HOST_RINGS 64 /* the adapter supports the host rings */
757 #define NAF_FORCE_NATIVE 128 /* the adapter is always NATIVE */
759 #define NAF_MOREFRAG 512 /* the adapter supports NS_MOREFRAG */
760 #define NAF_OFFSETS 1024 /* the adapter supports the slot offsets */
761 #define NAF_HOST_ALL 2048 /* the adapter wants as many host rings as hw */
762 #define NAF_ZOMBIE (1U<<30) /* the nic driver has been unloaded */
763 #define NAF_BUSY (1U<<31) /* the adapter is used internally and
764 * cannot be registered from userspace
766 int active_fds; /* number of user-space descriptors using this
767 interface, which is equal to the number of
768 struct netmap_if objs in the mapped region. */
770 u_int num_rx_rings; /* number of adapter receive rings */
771 u_int num_tx_rings; /* number of adapter transmit rings */
772 u_int num_host_rx_rings; /* number of host receive rings */
773 u_int num_host_tx_rings; /* number of host transmit rings */
775 u_int num_tx_desc; /* number of descriptor in each queue */
778 /* tx_rings and rx_rings are private but allocated as a
779 * contiguous chunk of memory. Each array has N+K entries,
780 * N for the hardware rings and K for the host rings.
782 struct netmap_kring **tx_rings; /* array of TX rings. */
783 struct netmap_kring **rx_rings; /* array of RX rings. */
785 void *tailroom; /* space below the rings array */
786 /* (used for leases) */
789 NM_SELINFO_T si[NR_TXRX]; /* global wait queues */
791 /* count users of the global wait queues */
792 int si_users[NR_TXRX];
794 void *pdev; /* used to store pci device */
796 /* copy of if_qflush and if_transmit pointers, to intercept
797 * packets from the network stack when netmap is active.
799 int (*if_transmit)(struct ifnet *, struct mbuf *);
801 /* copy of if_input for netmap_send_up() */
802 void (*if_input)(struct ifnet *, struct mbuf *);
804 /* Back reference to the parent ifnet struct. Used for
805 * hardware ports (emulated netmap included). */
806 struct ifnet *ifp; /* adapter is ifp->if_softc */
808 /*---- callbacks for this netmap adapter -----*/
810 * nm_dtor() is the cleanup routine called when destroying
812 * Called with NMG_LOCK held.
814 * nm_register() is called on NIOCREGIF and close() to enter
815 * or exit netmap mode on the NIC
816 * Called with NNG_LOCK held.
818 * nm_txsync() pushes packets to the underlying hw/switch
820 * nm_rxsync() collects packets from the underlying hw/switch
822 * nm_config() returns configuration information from the OS
823 * Called with NMG_LOCK held.
826 * the purpose of this callback is to fill the kring->hwbuf_len
827 * (l) and kring->buf_align fields. The l value is most important
828 * for RX rings, where we want to disallow writes outside of the
829 * netmap buffer. The l value must be computed taking into account
830 * the stipulated max_offset (o), possibly increased if there are
831 * alignment constraints, the maxframe (m), if known, and the
832 * current NETMAP_BUF_SIZE (b) of the memory region used by the
833 * adapter. We want the largest supported l such that o + l <= b.
834 * If m is known to be <= b - o, the callback may also choose the
835 * largest l <= m, ignoring the offset. The buf_align field is
836 * most important for TX rings when there are offsets. The user
837 * will see this value in the ring->buf_align field. Misaligned
838 * offsets will cause the corresponding packets to be silently
841 * nm_krings_create() create and init the tx_rings and
842 * rx_rings arrays of kring structures. In particular,
843 * set the nm_sync callbacks for each ring.
844 * There is no need to also allocate the corresponding
845 * netmap_rings, since netmap_mem_rings_create() will always
846 * be called to provide the missing ones.
847 * Called with NNG_LOCK held.
849 * nm_krings_delete() cleanup and delete the tx_rings and rx_rings
851 * Called with NMG_LOCK held.
853 * nm_notify() is used to act after data have become available
854 * (or the stopped state of the ring has changed)
855 * For hw devices this is typically a selwakeup(),
856 * but for NIC/host ports attached to a switch (or vice-versa)
857 * we also need to invoke the 'txsync' code downstream.
858 * This callback pointer is actually used only to initialize
860 * Return values are the same as for netmap_rx_irq().
862 void (*nm_dtor)(struct netmap_adapter *);
864 int (*nm_register)(struct netmap_adapter *, int onoff);
865 void (*nm_intr)(struct netmap_adapter *, int onoff);
867 int (*nm_txsync)(struct netmap_kring *kring, int flags);
868 int (*nm_rxsync)(struct netmap_kring *kring, int flags);
869 int (*nm_notify)(struct netmap_kring *kring, int flags);
870 int (*nm_bufcfg)(struct netmap_kring *kring, uint64_t target);
871 #define NAF_FORCE_READ 1
872 #define NAF_FORCE_RECLAIM 2
873 #define NAF_CAN_FORWARD_DOWN 4
874 /* return configuration information */
875 int (*nm_config)(struct netmap_adapter *, struct nm_config_info *info);
876 int (*nm_krings_create)(struct netmap_adapter *);
877 void (*nm_krings_delete)(struct netmap_adapter *);
879 * nm_bdg_attach() initializes the na_vp field to point
880 * to an adapter that can be attached to a VALE switch. If the
881 * current adapter is already a VALE port, na_vp is simply a cast;
882 * otherwise, na_vp points to a netmap_bwrap_adapter.
883 * If applicable, this callback also initializes na_hostvp,
884 * that can be used to connect the adapter host rings to the
886 * Called with NMG_LOCK held.
888 * nm_bdg_ctl() is called on the actual attach/detach to/from
889 * to/from the switch, to perform adapter-specific
891 * Called with NMG_LOCK held.
893 int (*nm_bdg_attach)(const char *bdg_name, struct netmap_adapter *,
895 int (*nm_bdg_ctl)(struct nmreq_header *, struct netmap_adapter *);
897 /* adapter used to attach this adapter to a VALE switch (if any) */
898 struct netmap_vp_adapter *na_vp;
899 /* adapter used to attach the host rings of this adapter
900 * to a VALE switch (if any) */
901 struct netmap_vp_adapter *na_hostvp;
903 /* standard refcount to control the lifetime of the adapter
904 * (it should be equal to the lifetime of the corresponding ifp)
908 /* memory allocator (opaque)
909 * We also cache a pointer to the lut_entry for translating
910 * buffer addresses, the total number of buffers and the buffer size.
912 struct netmap_mem_d *nm_mem;
913 struct netmap_mem_d *nm_mem_prev;
914 struct netmap_lut na_lut;
916 /* additional information attached to this adapter
917 * by other netmap subsystems. Currently used by
918 * bwrap, LINUX/v1000 and ptnetmap
922 /* array of pipes that have this adapter as a parent */
923 struct netmap_pipe_adapter **na_pipes;
924 int na_next_pipe; /* next free slot in the array */
925 int na_max_pipes; /* size of the array */
927 /* Offset of ethernet header for each packet. */
930 /* Max number of bytes that the NIC can store in the buffer
931 * referenced by each RX descriptor. This translates to the maximum
932 * bytes that a single netmap slot can reference. Larger packets
933 * require NS_MOREFRAG support. */
934 unsigned rx_buf_maxsize;
936 char name[NETMAP_REQ_IFNAMSIZ]; /* used at least by pipes */
939 unsigned long monitor_id; /* debugging */
943 static __inline u_int
944 nma_get_ndesc(struct netmap_adapter *na, enum txrx t)
946 return (t == NR_TX ? na->num_tx_desc : na->num_rx_desc);
950 nma_set_ndesc(struct netmap_adapter *na, enum txrx t, u_int v)
958 static __inline u_int
959 nma_get_nrings(struct netmap_adapter *na, enum txrx t)
961 return (t == NR_TX ? na->num_tx_rings : na->num_rx_rings);
964 static __inline u_int
965 nma_get_host_nrings(struct netmap_adapter *na, enum txrx t)
967 return (t == NR_TX ? na->num_host_tx_rings : na->num_host_rx_rings);
971 nma_set_nrings(struct netmap_adapter *na, enum txrx t, u_int v)
974 na->num_tx_rings = v;
976 na->num_rx_rings = v;
980 nma_set_host_nrings(struct netmap_adapter *na, enum txrx t, u_int v)
983 na->num_host_tx_rings = v;
985 na->num_host_rx_rings = v;
988 static __inline struct netmap_kring**
989 NMR(struct netmap_adapter *na, enum txrx t)
991 return (t == NR_TX ? na->tx_rings : na->rx_rings);
994 int nma_intr_enable(struct netmap_adapter *na, int onoff);
997 * If the NIC is owned by the kernel
998 * (i.e., bridge), neither another bridge nor user can use it;
999 * if the NIC is owned by a user, only users can share it.
1000 * Evaluation must be done under NMG_LOCK().
1002 #define NETMAP_OWNED_BY_KERN(na) ((na)->na_flags & NAF_BUSY)
1003 #define NETMAP_OWNED_BY_ANY(na) \
1004 (NETMAP_OWNED_BY_KERN(na) || ((na)->active_fds > 0))
1007 * derived netmap adapters for various types of ports
1009 struct netmap_vp_adapter { /* VALE software port */
1010 struct netmap_adapter up;
1015 * bdg_port is the port number used in the bridge;
1016 * na_bdg points to the bridge this NA is attached to.
1019 struct nm_bridge *na_bdg;
1021 int autodelete; /* remove the ifp on last reference */
1023 /* Maximum Frame Size, used in bdg_mismatch_datapath() */
1025 /* Last source MAC on this port */
1030 struct netmap_hw_adapter { /* physical device */
1031 struct netmap_adapter up;
1034 struct net_device_ops nm_ndo;
1035 struct ethtool_ops nm_eto;
1037 const struct ethtool_ops* save_ethtool;
1039 int (*nm_hw_register)(struct netmap_adapter *, int onoff);
1043 /* Mitigation support. */
1044 struct nm_generic_mit {
1045 struct hrtimer mit_timer;
1047 int mit_ring_idx; /* index of the ring being mitigated */
1048 struct netmap_adapter *mit_na; /* backpointer */
1051 struct netmap_generic_adapter { /* emulated device */
1052 struct netmap_hw_adapter up;
1054 /* Pointer to a previously used netmap adapter. */
1055 struct netmap_adapter *prev;
1057 /* Emulated netmap adapters support:
1058 * - save_if_input saves the if_input hook (FreeBSD);
1059 * - mit implements rx interrupt mitigation;
1061 void (*save_if_input)(struct ifnet *, struct mbuf *);
1063 struct nm_generic_mit *mit;
1065 netdev_tx_t (*save_start_xmit)(struct mbuf *, struct ifnet *);
1067 /* Is the adapter able to use multiple RX slots to scatter
1068 * each packet pushed up by the driver? */
1071 /* Is the transmission path controlled by a netmap-aware
1072 * device queue (i.e. qdisc on linux)? */
1075 #endif /* WITH_GENERIC */
1077 static __inline u_int
1078 netmap_real_rings(struct netmap_adapter *na, enum txrx t)
1080 return nma_get_nrings(na, t) +
1081 !!(na->na_flags & NAF_HOST_RINGS) * nma_get_host_nrings(na, t);
1084 /* account for fake rings */
1085 static __inline u_int
1086 netmap_all_rings(struct netmap_adapter *na, enum txrx t)
1088 return max(nma_get_nrings(na, t) + 1, netmap_real_rings(na, t));
1091 int netmap_default_bdg_attach(const char *name, struct netmap_adapter *na,
1092 struct nm_bridge *);
1093 struct nm_bdg_polling_state;
1095 * Bridge wrapper for non VALE ports attached to a VALE switch.
1097 * The real device must already have its own netmap adapter (hwna).
1098 * The bridge wrapper and the hwna adapter share the same set of
1099 * netmap rings and buffers, but they have two separate sets of
1100 * krings descriptors, with tx/rx meanings swapped:
1103 * bwrap krings rings krings hwna
1104 * +------+ +------+ +-----+ +------+ +------+
1105 * |tx_rings->| |\ /| |----| |<-tx_rings|
1106 * | | +------+ \ / +-----+ +------+ | |
1109 * | | +------+/ \+-----+ +------+ | |
1110 * |rx_rings->| | | |----| |<-rx_rings|
1111 * | | +------+ +-----+ +------+ | |
1114 * - packets coming from the bridge go to the brwap rx rings,
1115 * which are also the hwna tx rings. The bwrap notify callback
1116 * will then complete the hwna tx (see netmap_bwrap_notify).
1118 * - packets coming from the outside go to the hwna rx rings,
1119 * which are also the bwrap tx rings. The (overwritten) hwna
1120 * notify method will then complete the bridge tx
1121 * (see netmap_bwrap_intr_notify).
1123 * The bridge wrapper may optionally connect the hwna 'host' rings
1124 * to the bridge. This is done by using a second port in the
1125 * bridge and connecting it to the 'host' netmap_vp_adapter
1126 * contained in the netmap_bwrap_adapter. The brwap host adapter
1127 * cross-links the hwna host rings in the same way as shown above.
1129 * - packets coming from the bridge and directed to the host stack
1130 * are handled by the bwrap host notify callback
1131 * (see netmap_bwrap_host_notify)
1133 * - packets coming from the host stack are still handled by the
1134 * overwritten hwna notify callback (netmap_bwrap_intr_notify),
1135 * but are diverted to the host adapter depending on the ring number.
1138 struct netmap_bwrap_adapter {
1139 struct netmap_vp_adapter up;
1140 struct netmap_vp_adapter host; /* for host rings */
1141 struct netmap_adapter *hwna; /* the underlying device */
1144 * When we attach a physical interface to the bridge, we
1145 * allow the controlling process to terminate, so we need
1146 * a place to store the n_detmap_priv_d data structure.
1147 * This is only done when physical interfaces
1148 * are attached to a bridge.
1150 struct netmap_priv_d *na_kpriv;
1151 struct nm_bdg_polling_state *na_polling_state;
1152 /* we overwrite the hwna->na_vp pointer, so we save
1153 * here its original value, to be restored at detach
1155 struct netmap_vp_adapter *saved_na_vp;
1156 int (*nm_intr_notify)(struct netmap_kring *kring, int flags);
1158 int nm_bdg_polling(struct nmreq_header *hdr);
1160 int netmap_bdg_attach(struct nmreq_header *hdr, void *auth_token);
1161 int netmap_bdg_detach(struct nmreq_header *hdr, void *auth_token);
1163 int netmap_vale_list(struct nmreq_header *hdr);
1164 int netmap_vi_create(struct nmreq_header *hdr, int);
1165 int nm_vi_create(struct nmreq_header *);
1166 int nm_vi_destroy(const char *name);
1167 #else /* !WITH_VALE */
1168 #define netmap_vi_create(hdr, a) (EOPNOTSUPP)
1169 #endif /* WITH_VALE */
1173 #define NM_MAXPIPES 64 /* max number of pipes per adapter */
1175 struct netmap_pipe_adapter {
1176 /* pipe identifier is up.name */
1177 struct netmap_adapter up;
1179 #define NM_PIPE_ROLE_MASTER 0x1
1180 #define NM_PIPE_ROLE_SLAVE 0x2
1181 int role; /* either NM_PIPE_ROLE_MASTER or NM_PIPE_ROLE_SLAVE */
1183 struct netmap_adapter *parent; /* adapter that owns the memory */
1184 struct netmap_pipe_adapter *peer; /* the other end of the pipe */
1185 int peer_ref; /* 1 iff we are holding a ref to the peer */
1186 struct ifnet *parent_ifp; /* maybe null */
1188 u_int parent_slot; /* index in the parent pipe array */
1191 #endif /* WITH_PIPES */
1194 struct netmap_null_adapter {
1195 struct netmap_adapter up;
1197 #endif /* WITH_NMNULL */
1200 /* return slots reserved to rx clients; used in drivers */
1201 static inline uint32_t
1202 nm_kr_rxspace(struct netmap_kring *k)
1204 int space = k->nr_hwtail - k->nr_hwcur;
1206 space += k->nkr_num_slots;
1207 nm_prdis("preserving %d rx slots %d -> %d", space, k->nr_hwcur, k->nr_hwtail);
1212 /* return slots reserved to tx clients */
1213 #define nm_kr_txspace(_k) nm_kr_rxspace(_k)
1216 /* True if no space in the tx ring, only valid after txsync_prologue */
1218 nm_kr_txempty(struct netmap_kring *kring)
1220 return kring->rhead == kring->nr_hwtail;
1223 /* True if no more completed slots in the rx ring, only valid after
1224 * rxsync_prologue */
1225 #define nm_kr_rxempty(_k) nm_kr_txempty(_k)
1227 /* True if the application needs to wait for more space on the ring
1228 * (more received packets or more free tx slots).
1229 * Only valid after *xsync_prologue. */
1231 nm_kr_wouldblock(struct netmap_kring *kring)
1233 return kring->rcur == kring->nr_hwtail;
1237 * protect against multiple threads using the same ring.
1238 * also check that the ring has not been stopped or locked
1240 #define NM_KR_BUSY 1 /* some other thread is syncing the ring */
1241 #define NM_KR_STOPPED 2 /* unbounded stop (ifconfig down or driver unload) */
1242 #define NM_KR_LOCKED 3 /* bounded, brief stop for mutual exclusion */
1245 /* release the previously acquired right to use the *sync() methods of the ring */
1246 static __inline void nm_kr_put(struct netmap_kring *kr)
1248 NM_ATOMIC_CLEAR(&kr->nr_busy);
1252 /* true if the ifp that backed the adapter has disappeared (e.g., the
1253 * driver has been unloaded)
1255 static inline int nm_iszombie(struct netmap_adapter *na);
1257 /* try to obtain exclusive right to issue the *sync() operations on the ring.
1258 * The right is obtained and must be later relinquished via nm_kr_put() if and
1259 * only if nm_kr_tryget() returns 0.
1260 * If can_sleep is 1 there are only two other possible outcomes:
1261 * - the function returns NM_KR_BUSY
1262 * - the function returns NM_KR_STOPPED and sets the POLLERR bit in *perr
1264 * In both cases the caller will typically skip the ring, possibly collecting
1265 * errors along the way.
1266 * If the calling context does not allow sleeping, the caller must pass 0 in can_sleep.
1267 * In the latter case, the function may also return NM_KR_LOCKED and leave *perr
1268 * untouched: ideally, the caller should try again at a later time.
1270 static __inline int nm_kr_tryget(struct netmap_kring *kr, int can_sleep, int *perr)
1272 int busy = 1, stopped;
1273 /* check a first time without taking the lock
1274 * to avoid starvation for nm_kr_get()
1277 stopped = kr->nkr_stopped;
1278 if (unlikely(stopped)) {
1281 busy = NM_ATOMIC_TEST_AND_SET(&kr->nr_busy);
1282 /* we should not return NM_KR_BUSY if the ring was
1283 * actually stopped, so check another time after
1284 * the barrier provided by the atomic operation
1286 stopped = kr->nkr_stopped;
1287 if (unlikely(stopped)) {
1291 if (unlikely(nm_iszombie(kr->na))) {
1292 stopped = NM_KR_STOPPED;
1296 return unlikely(busy) ? NM_KR_BUSY : 0;
1301 if (stopped == NM_KR_STOPPED) {
1302 /* if POLLERR is defined we want to use it to simplify netmap_poll().
1303 * Otherwise, any non-zero value will do.
1306 #define NM_POLLERR POLLERR
1308 #define NM_POLLERR 1
1309 #endif /* POLLERR */
1311 *perr |= NM_POLLERR;
1313 } else if (can_sleep) {
1314 tsleep(kr, 0, "NM_KR_TRYGET", 4);
1320 /* put the ring in the 'stopped' state and wait for the current user (if any) to
1321 * notice. stopped must be either NM_KR_STOPPED or NM_KR_LOCKED
1323 static __inline void nm_kr_stop(struct netmap_kring *kr, int stopped)
1325 kr->nkr_stopped = stopped;
1326 while (NM_ATOMIC_TEST_AND_SET(&kr->nr_busy))
1327 tsleep(kr, 0, "NM_KR_GET", 4);
1330 /* restart a ring after a stop */
1331 static __inline void nm_kr_start(struct netmap_kring *kr)
1333 kr->nkr_stopped = 0;
1339 * The following functions are used by individual drivers to
1340 * support netmap operation.
1342 * netmap_attach() initializes a struct netmap_adapter, allocating the
1343 * struct netmap_ring's and the struct selinfo.
1345 * netmap_detach() frees the memory allocated by netmap_attach().
1347 * netmap_transmit() replaces the if_transmit routine of the interface,
1348 * and is used to intercept packets coming from the stack.
1350 * netmap_load_map/netmap_reload_map are helper routines to set/reset
1351 * the dmamap for a packet buffer
1353 * netmap_reset() is a helper routine to be called in the hw driver
1354 * when reinitializing a ring. It should not be called by
1355 * virtual ports (vale, pipes, monitor)
1357 int netmap_attach(struct netmap_adapter *);
1358 int netmap_attach_ext(struct netmap_adapter *, size_t size, int override_reg);
1359 void netmap_detach(struct ifnet *);
1360 int netmap_transmit(struct ifnet *, struct mbuf *);
1361 struct netmap_slot *netmap_reset(struct netmap_adapter *na,
1362 enum txrx tx, u_int n, u_int new_cur);
1363 int netmap_ring_reinit(struct netmap_kring *);
1364 int netmap_rings_config_get(struct netmap_adapter *, struct nm_config_info *);
1366 /* Return codes for netmap_*x_irq. */
1368 /* Driver should do normal interrupt processing, e.g. because
1369 * the interface is not in netmap mode. */
1371 /* Port is in netmap mode, and the interrupt work has been
1372 * completed. The driver does not have to notify netmap
1373 * again before the next interrupt. */
1374 NM_IRQ_COMPLETED = -1,
1375 /* Port is in netmap mode, but the interrupt work has not been
1376 * completed. The driver has to make sure netmap will be
1377 * notified again soon, even if no more interrupts come (e.g.
1378 * on Linux the driver should not call napi_complete()). */
1379 NM_IRQ_RESCHED = -2,
1382 /* default functions to handle rx/tx interrupts */
1383 int netmap_rx_irq(struct ifnet *, u_int, u_int *);
1384 #define netmap_tx_irq(_n, _q) netmap_rx_irq(_n, _q, NULL)
1385 int netmap_common_irq(struct netmap_adapter *, u_int, u_int *work_done);
1389 /* functions used by external modules to interface with VALE */
1390 #define netmap_vp_to_ifp(_vp) ((_vp)->up.ifp)
1391 #define netmap_ifp_to_vp(_ifp) (NA(_ifp)->na_vp)
1392 #define netmap_ifp_to_host_vp(_ifp) (NA(_ifp)->na_hostvp)
1393 #define netmap_bdg_idx(_vp) ((_vp)->bdg_port)
1394 const char *netmap_bdg_name(struct netmap_vp_adapter *);
1395 #else /* !WITH_VALE */
1396 #define netmap_vp_to_ifp(_vp) NULL
1397 #define netmap_ifp_to_vp(_ifp) NULL
1398 #define netmap_ifp_to_host_vp(_ifp) NULL
1399 #define netmap_bdg_idx(_vp) -1
1400 #endif /* WITH_VALE */
1403 nm_netmap_on(struct netmap_adapter *na)
1405 return na && na->na_flags & NAF_NETMAP_ON;
1409 nm_native_on(struct netmap_adapter *na)
1411 return nm_netmap_on(na) && (na->na_flags & NAF_NATIVE);
1414 static inline struct netmap_kring *
1415 netmap_kring_on(struct netmap_adapter *na, u_int q, enum txrx t)
1417 struct netmap_kring *kring = NULL;
1419 if (!nm_native_on(na))
1422 if (t == NR_RX && q < na->num_rx_rings)
1423 kring = na->rx_rings[q];
1424 else if (t == NR_TX && q < na->num_tx_rings)
1425 kring = na->tx_rings[q];
1429 return (kring->nr_mode == NKR_NETMAP_ON) ? kring : NULL;
1433 nm_iszombie(struct netmap_adapter *na)
1435 return na == NULL || (na->na_flags & NAF_ZOMBIE);
1438 void nm_set_native_flags(struct netmap_adapter *);
1439 void nm_clear_native_flags(struct netmap_adapter *);
1441 void netmap_krings_mode_commit(struct netmap_adapter *na, int onoff);
1444 * nm_*sync_prologue() functions are used in ioctl/poll and ptnetmap
1446 * We need netmap_ring* parameter, because in ptnetmap it is decoupled
1448 * The user-space ring pointers (head/cur/tail) are shared through
1449 * CSB between host and guest.
1453 * validates parameters in the ring/kring, returns a value for head
1454 * If any error, returns ring_size to force a reinit.
1456 uint32_t nm_txsync_prologue(struct netmap_kring *, struct netmap_ring *);
1460 * validates parameters in the ring/kring, returns a value for head
1461 * If any error, returns ring_size lim to force a reinit.
1463 uint32_t nm_rxsync_prologue(struct netmap_kring *, struct netmap_ring *);
1466 /* check/fix address and len in tx rings */
1467 #if 1 /* debug version */
1468 #define NM_CHECK_ADDR_LEN(_na, _a, _l) do { \
1469 if (_a == NETMAP_BUF_BASE(_na) || _l > NETMAP_BUF_SIZE(_na)) { \
1470 nm_prlim(5, "bad addr/len ring %d slot %d idx %d len %d", \
1471 kring->ring_id, nm_i, slot->buf_idx, len); \
1472 if (_l > NETMAP_BUF_SIZE(_na)) \
1473 _l = NETMAP_BUF_SIZE(_na); \
1475 #else /* no debug version */
1476 #define NM_CHECK_ADDR_LEN(_na, _a, _l) do { \
1477 if (_l > NETMAP_BUF_SIZE(_na)) \
1478 _l = NETMAP_BUF_SIZE(_na); \
1482 #define NM_CHECK_ADDR_LEN_OFF(na_, l_, o_) do { \
1483 if ((l_) + (o_) < (l_) || \
1484 (l_) + (o_) > NETMAP_BUF_SIZE(na_)) { \
1485 (l_) = NETMAP_BUF_SIZE(na_) - (o_); \
1489 /*---------------------------------------------------------------*/
1491 * Support routines used by netmap subsystems
1492 * (native drivers, VALE, generic, pipes, monitors, ...)
1496 /* common routine for all functions that create a netmap adapter. It performs
1498 * - if the na points to an ifp, mark the ifp as netmap capable
1499 * using na as its native adapter;
1500 * - provide defaults for the setup callbacks and the memory allocator
1502 int netmap_attach_common(struct netmap_adapter *);
1503 /* fill priv->np_[tr]xq{first,last} using the ringid and flags information
1504 * coming from a struct nmreq_register
1506 int netmap_interp_ringid(struct netmap_priv_d *priv, struct nmreq_header *hdr);
1507 /* update the ring parameters (number and size of tx and rx rings).
1508 * It calls the nm_config callback, if available.
1510 int netmap_update_config(struct netmap_adapter *na);
1511 /* create and initialize the common fields of the krings array.
1512 * using the information that must be already available in the na.
1513 * tailroom can be used to request the allocation of additional
1514 * tailroom bytes after the krings array. This is used by
1515 * netmap_vp_adapter's (i.e., VALE ports) to make room for
1516 * leasing-related data structures
1518 int netmap_krings_create(struct netmap_adapter *na, u_int tailroom);
1519 /* deletes the kring array of the adapter. The array must have
1520 * been created using netmap_krings_create
1522 void netmap_krings_delete(struct netmap_adapter *na);
1524 int netmap_hw_krings_create(struct netmap_adapter *na);
1525 void netmap_hw_krings_delete(struct netmap_adapter *na);
1527 /* set the stopped/enabled status of ring
1528 * When stopping, they also wait for all current activity on the ring to
1529 * terminate. The status change is then notified using the na nm_notify
1532 void netmap_set_ring(struct netmap_adapter *, u_int ring_id, enum txrx, int stopped);
1533 /* set the stopped/enabled status of all rings of the adapter. */
1534 void netmap_set_all_rings(struct netmap_adapter *, int stopped);
1535 /* convenience wrappers for netmap_set_all_rings */
1536 void netmap_disable_all_rings(struct ifnet *);
1537 void netmap_enable_all_rings(struct ifnet *);
1539 int netmap_buf_size_validate(const struct netmap_adapter *na, unsigned mtu);
1540 int netmap_do_regif(struct netmap_priv_d *priv, struct netmap_adapter *na,
1541 struct nmreq_header *);
1542 void netmap_do_unregif(struct netmap_priv_d *priv);
1544 u_int nm_bound_var(u_int *v, u_int dflt, u_int lo, u_int hi, const char *msg);
1545 int netmap_get_na(struct nmreq_header *hdr, struct netmap_adapter **na,
1546 struct ifnet **ifp, struct netmap_mem_d *nmd, int create);
1547 void netmap_unget_na(struct netmap_adapter *na, struct ifnet *ifp);
1548 int netmap_get_hw_na(struct ifnet *ifp,
1549 struct netmap_mem_d *nmd, struct netmap_adapter **na);
1550 void netmap_mem_restore(struct netmap_adapter *na);
1553 uint32_t netmap_vale_learning(struct nm_bdg_fwd *ft, uint8_t *dst_ring,
1554 struct netmap_vp_adapter *, void *private_data);
1556 /* these are redefined in case of no VALE support */
1557 int netmap_get_vale_na(struct nmreq_header *hdr, struct netmap_adapter **na,
1558 struct netmap_mem_d *nmd, int create);
1559 void *netmap_vale_create(const char *bdg_name, int *return_status);
1560 int netmap_vale_destroy(const char *bdg_name, void *auth_token);
1562 #else /* !WITH_VALE */
1563 #define netmap_bdg_learning(_1, _2, _3, _4) 0
1564 #define netmap_get_vale_na(_1, _2, _3, _4) 0
1565 #define netmap_bdg_create(_1, _2) NULL
1566 #define netmap_bdg_destroy(_1, _2) 0
1567 #endif /* !WITH_VALE */
1570 /* max number of pipes per device */
1571 #define NM_MAXPIPES 64 /* XXX this should probably be a sysctl */
1572 void netmap_pipe_dealloc(struct netmap_adapter *);
1573 int netmap_get_pipe_na(struct nmreq_header *hdr, struct netmap_adapter **na,
1574 struct netmap_mem_d *nmd, int create);
1575 #else /* !WITH_PIPES */
1576 #define NM_MAXPIPES 0
1577 #define netmap_pipe_alloc(_1, _2) 0
1578 #define netmap_pipe_dealloc(_1)
1579 #define netmap_get_pipe_na(hdr, _2, _3, _4) \
1580 ((strchr(hdr->nr_name, '{') != NULL || strchr(hdr->nr_name, '}') != NULL) ? EOPNOTSUPP : 0)
1584 int netmap_get_monitor_na(struct nmreq_header *hdr, struct netmap_adapter **na,
1585 struct netmap_mem_d *nmd, int create);
1586 void netmap_monitor_stop(struct netmap_adapter *na);
1588 #define netmap_get_monitor_na(hdr, _2, _3, _4) \
1589 (((struct nmreq_register *)(uintptr_t)hdr->nr_body)->nr_flags & (NR_MONITOR_TX | NR_MONITOR_RX) ? EOPNOTSUPP : 0)
1593 int netmap_get_null_na(struct nmreq_header *hdr, struct netmap_adapter **na,
1594 struct netmap_mem_d *nmd, int create);
1595 #else /* !WITH_NMNULL */
1596 #define netmap_get_null_na(hdr, _2, _3, _4) \
1597 (((struct nmreq_register *)(uintptr_t)hdr->nr_body)->nr_flags & (NR_MONITOR_TX | NR_MONITOR_RX) ? EOPNOTSUPP : 0)
1598 #endif /* WITH_NMNULL */
1600 #ifdef CONFIG_NET_NS
1601 struct net *netmap_bns_get(void);
1602 void netmap_bns_put(struct net *);
1603 void netmap_bns_getbridges(struct nm_bridge **, u_int *);
1605 extern struct nm_bridge *nm_bridges;
1606 #define netmap_bns_get()
1607 #define netmap_bns_put(_1)
1608 #define netmap_bns_getbridges(b, n) \
1609 do { *b = nm_bridges; *n = NM_BRIDGES; } while (0)
1612 /* Various prototypes */
1613 int netmap_poll(struct netmap_priv_d *, int events, NM_SELRECORD_T *td);
1614 int netmap_init(void);
1615 void netmap_fini(void);
1616 int netmap_get_memory(struct netmap_priv_d* p);
1617 void netmap_dtor(void *data);
1619 int netmap_ioctl(struct netmap_priv_d *priv, u_long cmd, caddr_t data,
1620 struct thread *, int nr_body_is_user);
1621 int netmap_ioctl_legacy(struct netmap_priv_d *priv, u_long cmd, caddr_t data,
1623 size_t nmreq_size_by_type(uint16_t nr_reqtype);
1625 /* netmap_adapter creation/destruction */
1627 // #define NM_DEBUG_PUTGET 1
1629 #ifdef NM_DEBUG_PUTGET
1631 #define NM_DBG(f) __##f
1633 void __netmap_adapter_get(struct netmap_adapter *na);
1635 #define netmap_adapter_get(na) \
1637 struct netmap_adapter *__na = na; \
1638 nm_prinf("getting %p:%s (%d)", __na, (__na)->name, (__na)->na_refcount); \
1639 __netmap_adapter_get(__na); \
1642 int __netmap_adapter_put(struct netmap_adapter *na);
1644 #define netmap_adapter_put(na) \
1646 struct netmap_adapter *__na = na; \
1647 nm_prinf("putting %p:%s (%d)", __na, (__na)->name, (__na)->na_refcount); \
1648 __netmap_adapter_put(__na); \
1651 #else /* !NM_DEBUG_PUTGET */
1654 void netmap_adapter_get(struct netmap_adapter *na);
1655 int netmap_adapter_put(struct netmap_adapter *na);
1657 #endif /* !NM_DEBUG_PUTGET */
1663 #define NETMAP_BUF_BASE(_na) ((_na)->na_lut.lut[0].vaddr)
1664 #define NETMAP_BUF_SIZE(_na) ((_na)->na_lut.objsize)
1665 extern int netmap_no_pendintr;
1666 extern int netmap_verbose;
1667 #ifdef CONFIG_NETMAP_DEBUG
1668 extern int netmap_debug; /* for debugging */
1669 #else /* !CONFIG_NETMAP_DEBUG */
1670 #define netmap_debug (0)
1671 #endif /* !CONFIG_NETMAP_DEBUG */
1672 enum { /* debug flags */
1673 NM_DEBUG_ON = 1, /* generic debug messages */
1674 NM_DEBUG_HOST = 0x2, /* debug host stack */
1675 NM_DEBUG_RXSYNC = 0x10, /* debug on rxsync/txsync */
1676 NM_DEBUG_TXSYNC = 0x20,
1677 NM_DEBUG_RXINTR = 0x100, /* debug on rx/tx intr (driver) */
1678 NM_DEBUG_TXINTR = 0x200,
1679 NM_DEBUG_NIC_RXSYNC = 0x1000, /* debug on rx/tx intr (driver) */
1680 NM_DEBUG_NIC_TXSYNC = 0x2000,
1681 NM_DEBUG_MEM = 0x4000, /* verbose memory allocations/deallocations */
1682 NM_DEBUG_VALE = 0x8000, /* debug messages from memory allocators */
1683 NM_DEBUG_BDG = NM_DEBUG_VALE,
1686 extern int netmap_txsync_retry;
1687 extern int netmap_generic_hwcsum;
1688 extern int netmap_generic_mit;
1689 extern int netmap_generic_ringsize;
1690 extern int netmap_generic_rings;
1692 extern int netmap_generic_txqdisc;
1696 * NA returns a pointer to the struct netmap adapter from the ifp.
1697 * WNA is os-specific and must be defined in glue code.
1699 #define NA(_ifp) ((struct netmap_adapter *)WNA(_ifp))
1702 * we provide a default implementation of NM_ATTACH_NA/NM_DETACH_NA
1703 * based on the WNA field.
1704 * Glue code may override this by defining its own NM_ATTACH_NA
1706 #ifndef NM_ATTACH_NA
1708 * On old versions of FreeBSD, NA(ifp) is a pspare. On linux we
1709 * overload another pointer in the netdev.
1711 * We check if NA(ifp) is set and its first element has a related
1712 * magic value. The capenable is within the struct netmap_adapter.
1714 #define NETMAP_MAGIC 0x52697a7a
1716 #define NM_NA_VALID(ifp) (NA(ifp) && \
1717 ((uint32_t)(uintptr_t)NA(ifp) ^ NA(ifp)->magic) == NETMAP_MAGIC )
1719 #define NM_ATTACH_NA(ifp, na) do { \
1723 ((uint32_t)(uintptr_t)NA(ifp)) ^ NETMAP_MAGIC; \
1725 #define NM_RESTORE_NA(ifp, na) WNA(ifp) = na;
1727 #define NM_DETACH_NA(ifp) do { WNA(ifp) = NULL; } while (0)
1728 #define NM_NA_CLASH(ifp) (NA(ifp) && !NM_NA_VALID(ifp))
1729 #endif /* !NM_ATTACH_NA */
1732 #define NM_IS_NATIVE(ifp) (NM_NA_VALID(ifp) && NA(ifp)->nm_dtor == netmap_hw_dtor)
1734 #if defined(__FreeBSD__)
1736 /* Assigns the device IOMMU domain to an allocator.
1737 * Returns -ENOMEM in case the domain is different */
1738 #define nm_iommu_group_id(dev) (-1)
1740 /* Callback invoked by the dma machinery after a successful dmamap_load */
1741 static void netmap_dmamap_cb(__unused void *arg,
1742 __unused bus_dma_segment_t * segs, __unused int nseg, __unused int error)
1746 /* bus_dmamap_load wrapper: call aforementioned function if map != NULL.
1747 * XXX can we do it without a callback ?
1750 netmap_load_map(struct netmap_adapter *na,
1751 bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
1754 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE(na),
1755 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT);
1760 netmap_unload_map(struct netmap_adapter *na,
1761 bus_dma_tag_t tag, bus_dmamap_t map)
1764 bus_dmamap_unload(tag, map);
1767 #define netmap_sync_map(na, tag, map, sz, t)
1769 /* update the map when a buffer changes. */
1771 netmap_reload_map(struct netmap_adapter *na,
1772 bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
1775 bus_dmamap_unload(tag, map);
1776 bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE(na),
1777 netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT);
1781 #elif defined(_WIN32)
1785 int nm_iommu_group_id(bus_dma_tag_t dev);
1786 #include <linux/dma-mapping.h>
1790 * dma_map_single(&pdev->dev, virt_addr, len, direction)
1791 * dma_unmap_single(&adapter->pdev->dev, phys_addr, len, direction)
1794 struct e1000_buffer *buffer_info = &tx_ring->buffer_info[l];
1795 /* set time_stamp *before* dma to help avoid a possible race */
1796 buffer_info->time_stamp = jiffies;
1797 buffer_info->mapped_as_page = false;
1798 buffer_info->length = len;
1799 //buffer_info->next_to_watch = l;
1800 /* reload dma map */
1801 dma_unmap_single(&adapter->pdev->dev, buffer_info->dma,
1802 NETMAP_BUF_SIZE, DMA_TO_DEVICE);
1803 buffer_info->dma = dma_map_single(&adapter->pdev->dev,
1804 addr, NETMAP_BUF_SIZE, DMA_TO_DEVICE);
1806 if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) {
1807 nm_prerr("dma mapping error");
1808 /* goto dma_error; See e1000_put_txbuf() */
1811 tx_desc->buffer_addr = htole64(buffer_info->dma); //XXX
1816 netmap_load_map(struct netmap_adapter *na,
1817 bus_dma_tag_t tag, bus_dmamap_t map, void *buf, u_int size)
1820 *map = dma_map_single(na->pdev, buf, size,
1822 if (dma_mapping_error(na->pdev, *map)) {
1831 netmap_unload_map(struct netmap_adapter *na,
1832 bus_dma_tag_t tag, bus_dmamap_t map, u_int sz)
1835 dma_unmap_single(na->pdev, *map, sz,
1840 #ifdef NETMAP_LINUX_HAVE_DMASYNC
1842 netmap_sync_map_cpu(struct netmap_adapter *na,
1843 bus_dma_tag_t tag, bus_dmamap_t map, u_int sz, enum txrx t)
1846 dma_sync_single_for_cpu(na->pdev, *map, sz,
1847 (t == NR_TX ? DMA_TO_DEVICE : DMA_FROM_DEVICE));
1852 netmap_sync_map_dev(struct netmap_adapter *na,
1853 bus_dma_tag_t tag, bus_dmamap_t map, u_int sz, enum txrx t)
1856 dma_sync_single_for_device(na->pdev, *map, sz,
1857 (t == NR_TX ? DMA_TO_DEVICE : DMA_FROM_DEVICE));
1862 netmap_reload_map(struct netmap_adapter *na,
1863 bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
1865 u_int sz = NETMAP_BUF_SIZE(na);
1868 dma_unmap_single(na->pdev, *map, sz,
1872 *map = dma_map_single(na->pdev, buf, sz,
1875 #else /* !NETMAP_LINUX_HAVE_DMASYNC */
1876 #define netmap_sync_map_cpu(na, tag, map, sz, t)
1877 #define netmap_sync_map_dev(na, tag, map, sz, t)
1878 #endif /* NETMAP_LINUX_HAVE_DMASYNC */
1884 * functions to map NIC to KRING indexes (n2k) and vice versa (k2n)
1887 netmap_idx_n2k(struct netmap_kring *kr, int idx)
1889 int n = kr->nkr_num_slots;
1891 if (likely(kr->nkr_hwofs == 0)) {
1895 idx += kr->nkr_hwofs;
1906 netmap_idx_k2n(struct netmap_kring *kr, int idx)
1908 int n = kr->nkr_num_slots;
1910 if (likely(kr->nkr_hwofs == 0)) {
1914 idx -= kr->nkr_hwofs;
1924 /* Entries of the look-up table. */
1927 void *vaddr; /* virtual address. */
1928 vm_paddr_t paddr; /* physical address. */
1930 #else /* linux & _WIN32 */
1931 /* dma-mapping in linux can assign a buffer a different address
1932 * depending on the device, so we need to have a separate
1933 * physical-address look-up table for each na.
1934 * We can still share the vaddrs, though, therefore we split
1935 * the lut_entry structure.
1938 void *vaddr; /* virtual address. */
1942 vm_paddr_t paddr; /* physical address. */
1944 #endif /* linux & _WIN32 */
1946 struct netmap_obj_pool;
1948 /* alignment for netmap buffers */
1949 #define NM_BUF_ALIGN 64
1952 * NMB return the virtual address of a buffer (buffer 0 on bad index)
1953 * PNMB also fills the physical address
1955 static inline void *
1956 NMB(struct netmap_adapter *na, struct netmap_slot *slot)
1958 struct lut_entry *lut = na->na_lut.lut;
1959 uint32_t i = slot->buf_idx;
1960 return (unlikely(i >= na->na_lut.objtotal)) ?
1961 lut[0].vaddr : lut[i].vaddr;
1964 static inline void *
1965 PNMB(struct netmap_adapter *na, struct netmap_slot *slot, uint64_t *pp)
1967 uint32_t i = slot->buf_idx;
1968 struct lut_entry *lut = na->na_lut.lut;
1969 struct plut_entry *plut = na->na_lut.plut;
1970 void *ret = (i >= na->na_lut.objtotal) ? lut[0].vaddr : lut[i].vaddr;
1973 *pp = (i >= na->na_lut.objtotal) ? (uint64_t)plut[0].paddr.QuadPart : (uint64_t)plut[i].paddr.QuadPart;
1975 *pp = (i >= na->na_lut.objtotal) ? plut[0].paddr : plut[i].paddr;
1981 nm_write_offset(struct netmap_kring *kring,
1982 struct netmap_slot *slot, uint64_t offset)
1984 slot->ptr = (slot->ptr & ~kring->offset_mask) |
1985 (offset & kring->offset_mask);
1988 static inline uint64_t
1989 nm_get_offset(struct netmap_kring *kring, struct netmap_slot *slot)
1991 uint64_t offset = (slot->ptr & kring->offset_mask);
1992 if (unlikely(offset > kring->offset_max))
1993 offset = kring->offset_max;
1997 static inline void *
1998 NMB_O(struct netmap_kring *kring, struct netmap_slot *slot)
2000 void *addr = NMB(kring->na, slot);
2001 return (char *)addr + nm_get_offset(kring, slot);
2004 static inline void *
2005 PNMB_O(struct netmap_kring *kring, struct netmap_slot *slot, uint64_t *pp)
2007 void *addr = PNMB(kring->na, slot, pp);
2008 uint64_t offset = nm_get_offset(kring, slot);
2009 addr = (char *)addr + offset;
2016 * Structure associated to each netmap file descriptor.
2017 * It is created on open and left unbound (np_nifp == NULL).
2018 * A successful NIOCREGIF will set np_nifp and the first few fields;
2019 * this is protected by a global lock (NMG_LOCK) due to low contention.
2021 * np_refs counts the number of references to the structure: one for the fd,
2022 * plus (on FreeBSD) one for each active mmap which we track ourselves
2023 * (linux automatically tracks them, but FreeBSD does not).
2024 * np_refs is protected by NMG_LOCK.
2026 * Read access to the structure is lock free, because ni_nifp once set
2027 * can only go to 0 when nobody is using the entry anymore. Readers
2028 * must check that np_nifp != NULL before using the other fields.
2030 struct netmap_priv_d {
2031 struct netmap_if * volatile np_nifp; /* netmap if descriptor. */
2033 struct netmap_adapter *np_na;
2034 struct ifnet *np_ifp;
2035 uint32_t np_flags; /* from the ioctl */
2036 u_int np_qfirst[NR_TXRX],
2037 np_qlast[NR_TXRX]; /* range of tx/rx rings to scan */
2039 uint16_t np_kloop_state; /* use with NMG_LOCK held */
2040 #define NM_SYNC_KLOOP_RUNNING (1 << 0)
2041 #define NM_SYNC_KLOOP_STOPPING (1 << 1)
2042 int np_sync_flags; /* to be passed to nm_sync */
2044 int np_refs; /* use with NMG_LOCK held */
2046 /* pointers to the selinfo to be used for selrecord.
2047 * Either the local or the global one depending on the
2050 NM_SELINFO_T *np_si[NR_TXRX];
2052 /* In the optional CSB mode, the user must specify the start address
2053 * of two arrays of Communication Status Block (CSB) entries, for the
2054 * two directions (kernel read application write, and kernel write
2055 * application read).
2056 * The number of entries must agree with the number of rings bound to
2057 * the netmap file descriptor. The entries corresponding to the TX
2058 * rings are laid out before the ones corresponding to the RX rings.
2060 * Array of CSB entries for application --> kernel communication
2062 struct nm_csb_atok *np_csb_atok_base;
2063 /* Array of CSB entries for kernel --> application communication
2065 struct nm_csb_ktoa *np_csb_ktoa_base;
2068 struct file *np_filp; /* used by sync kloop */
2072 struct netmap_priv_d *netmap_priv_new(void);
2073 void netmap_priv_delete(struct netmap_priv_d *);
2075 static inline int nm_kring_pending(struct netmap_priv_d *np)
2077 struct netmap_adapter *na = np->np_na;
2082 for (i = np->np_qfirst[t]; i < np->np_qlast[t]; i++) {
2083 struct netmap_kring *kring = NMR(na, t)[i];
2084 if (kring->nr_mode != kring->nr_pending_mode) {
2092 /* call with NMG_LOCK held */
2094 nm_si_user(struct netmap_priv_d *priv, enum txrx t)
2096 return (priv->np_na != NULL &&
2097 (priv->np_qlast[t] - priv->np_qfirst[t] > 1));
2101 int netmap_pipe_txsync(struct netmap_kring *txkring, int flags);
2102 int netmap_pipe_rxsync(struct netmap_kring *rxkring, int flags);
2103 int netmap_pipe_krings_create_both(struct netmap_adapter *na,
2104 struct netmap_adapter *ona);
2105 void netmap_pipe_krings_delete_both(struct netmap_adapter *na,
2106 struct netmap_adapter *ona);
2107 int netmap_pipe_reg_both(struct netmap_adapter *na,
2108 struct netmap_adapter *ona);
2109 #endif /* WITH_PIPES */
2113 struct netmap_monitor_adapter {
2114 struct netmap_adapter up;
2116 struct netmap_priv_d priv;
2120 #endif /* WITH_MONITOR */
2125 * generic netmap emulation for devices that do not have
2126 * native netmap support.
2128 int generic_netmap_attach(struct ifnet *ifp);
2129 int generic_rx_handler(struct ifnet *ifp, struct mbuf *m);;
2131 int nm_os_catch_rx(struct netmap_generic_adapter *gna, int intercept);
2132 int nm_os_catch_tx(struct netmap_generic_adapter *gna, int intercept);
2134 int na_is_generic(struct netmap_adapter *na);
2137 * the generic transmit routine is passed a structure to optionally
2138 * build a queue of descriptors, in an OS-specific way.
2139 * The payload is at addr, if non-null, and the routine should send or queue
2140 * the packet, returning 0 if successful, 1 on failure.
2142 * At the end, if head is non-null, there will be an additional call
2143 * to the function with addr = NULL; this should tell the OS-specific
2144 * routine to send the queue and free any resources. Failure is ignored.
2146 struct nm_os_gen_arg {
2148 void *m; /* os-specific mbuf-like object */
2149 void *head, *tail; /* tailq, if the OS-specific routine needs to build one */
2150 void *addr; /* payload of current packet */
2151 u_int len; /* packet length */
2152 u_int ring_nr; /* packet length */
2153 u_int qevent; /* in txqdisc mode, place an event on this mbuf */
2156 int nm_os_generic_xmit_frame(struct nm_os_gen_arg *);
2157 int nm_os_generic_find_num_desc(struct ifnet *ifp, u_int *tx, u_int *rx);
2158 void nm_os_generic_find_num_queues(struct ifnet *ifp, u_int *txq, u_int *rxq);
2159 void nm_os_generic_set_features(struct netmap_generic_adapter *gna);
2161 static inline struct ifnet*
2162 netmap_generic_getifp(struct netmap_generic_adapter *gna)
2165 return gna->prev->ifp;
2167 return gna->up.up.ifp;
2170 void netmap_generic_irq(struct netmap_adapter *na, u_int q, u_int *work_done);
2172 //#define RATE_GENERIC /* Enables communication statistics for generic. */
2174 void generic_rate(int txp, int txs, int txi, int rxp, int rxs, int rxi);
2176 #define generic_rate(txp, txs, txi, rxp, rxs, rxi)
2180 * netmap_mitigation API. This is used by the generic adapter
2181 * to reduce the number of interrupt requests/selwakeup
2182 * to clients on incoming packets.
2184 void nm_os_mitigation_init(struct nm_generic_mit *mit, int idx,
2185 struct netmap_adapter *na);
2186 void nm_os_mitigation_start(struct nm_generic_mit *mit);
2187 void nm_os_mitigation_restart(struct nm_generic_mit *mit);
2188 int nm_os_mitigation_active(struct nm_generic_mit *mit);
2189 void nm_os_mitigation_cleanup(struct nm_generic_mit *mit);
2190 #else /* !WITH_GENERIC */
2191 #define generic_netmap_attach(ifp) (EOPNOTSUPP)
2192 #define na_is_generic(na) (0)
2193 #endif /* WITH_GENERIC */
2195 /* Shared declarations for the VALE switch. */
2198 * Each transmit queue accumulates a batch of packets into
2199 * a structure before forwarding. Packets to the same
2200 * destination are put in a list using ft_next as a link field.
2201 * ft_frags and ft_next are valid only on the first fragment.
2203 struct nm_bdg_fwd { /* forwarding entry for a bridge */
2204 void *ft_buf; /* netmap or indirect buffer */
2205 uint8_t ft_frags; /* how many fragments (only on 1st frag) */
2206 uint16_t ft_offset; /* dst port (unused) */
2207 uint16_t ft_flags; /* flags, e.g. indirect */
2208 uint16_t ft_len; /* src fragment len */
2209 uint16_t ft_next; /* next packet to same destination */
2212 /* struct 'virtio_net_hdr' from linux. */
2213 struct nm_vnet_hdr {
2214 #define VIRTIO_NET_HDR_F_NEEDS_CSUM 1 /* Use csum_start, csum_offset */
2215 #define VIRTIO_NET_HDR_F_DATA_VALID 2 /* Csum is valid */
2217 #define VIRTIO_NET_HDR_GSO_NONE 0 /* Not a GSO frame */
2218 #define VIRTIO_NET_HDR_GSO_TCPV4 1 /* GSO frame, IPv4 TCP (TSO) */
2219 #define VIRTIO_NET_HDR_GSO_UDP 3 /* GSO frame, IPv4 UDP (UFO) */
2220 #define VIRTIO_NET_HDR_GSO_TCPV6 4 /* GSO frame, IPv6 TCP */
2221 #define VIRTIO_NET_HDR_GSO_ECN 0x80 /* TCP has ECN set */
2225 uint16_t csum_start;
2226 uint16_t csum_offset;
2229 #define WORST_CASE_GSO_HEADER (14+40+60) /* IPv6 + TCP */
2231 /* Private definitions for IPv4, IPv6, UDP and TCP headers. */
2234 uint8_t version_ihl;
2244 /*The options start here. */
2252 uint8_t doff; /* Data offset + Reserved */
2267 uint8_t priority_version;
2268 uint8_t flow_lbl[3];
2270 uint16_t payload_len;
2278 /* Type used to store a checksum (in host byte order) that hasn't been
2281 #define rawsum_t uint32_t
2283 rawsum_t nm_os_csum_raw(uint8_t *data, size_t len, rawsum_t cur_sum);
2284 uint16_t nm_os_csum_ipv4(struct nm_iphdr *iph);
2285 void nm_os_csum_tcpudp_ipv4(struct nm_iphdr *iph, void *data,
2286 size_t datalen, uint16_t *check);
2287 void nm_os_csum_tcpudp_ipv6(struct nm_ipv6hdr *ip6h, void *data,
2288 size_t datalen, uint16_t *check);
2289 uint16_t nm_os_csum_fold(rawsum_t cur_sum);
2291 void bdg_mismatch_datapath(struct netmap_vp_adapter *na,
2292 struct netmap_vp_adapter *dst_na,
2293 const struct nm_bdg_fwd *ft_p,
2294 struct netmap_ring *dst_ring,
2295 u_int *j, u_int lim, u_int *howmany);
2297 /* persistent virtual port routines */
2298 int nm_os_vi_persist(const char *, struct ifnet **);
2299 void nm_os_vi_detach(struct ifnet *);
2300 void nm_os_vi_init_index(void);
2303 * kernel thread routines
2305 struct nm_kctx; /* OS-specific kernel context - opaque */
2306 typedef void (*nm_kctx_worker_fn_t)(void *data);
2308 /* kthread configuration */
2309 struct nm_kctx_cfg {
2310 long type; /* kthread type/identifier */
2311 nm_kctx_worker_fn_t worker_fn; /* worker function */
2312 void *worker_private;/* worker parameter */
2313 int attach_user; /* attach kthread to user process */
2315 /* kthread configuration */
2316 struct nm_kctx *nm_os_kctx_create(struct nm_kctx_cfg *cfg,
2318 int nm_os_kctx_worker_start(struct nm_kctx *);
2319 void nm_os_kctx_worker_stop(struct nm_kctx *);
2320 void nm_os_kctx_destroy(struct nm_kctx *);
2321 void nm_os_kctx_worker_setaff(struct nm_kctx *, int);
2322 u_int nm_os_ncpus(void);
2324 int netmap_sync_kloop(struct netmap_priv_d *priv,
2325 struct nmreq_header *hdr);
2326 int netmap_sync_kloop_stop(struct netmap_priv_d *priv);
2328 #ifdef WITH_PTNETMAP
2329 /* ptnetmap guest routines */
2332 * ptnetmap_memdev routines used to talk with ptnetmap_memdev device driver
2334 struct ptnetmap_memdev;
2335 int nm_os_pt_memdev_iomap(struct ptnetmap_memdev *, vm_paddr_t *, void **,
2337 void nm_os_pt_memdev_iounmap(struct ptnetmap_memdev *);
2338 uint32_t nm_os_pt_memdev_ioread(struct ptnetmap_memdev *, unsigned int);
2341 * netmap adapter for guest ptnetmap ports
2343 struct netmap_pt_guest_adapter {
2344 /* The netmap adapter to be used by netmap applications.
2345 * This field must be the first, to allow upcast. */
2346 struct netmap_hw_adapter hwup;
2348 /* The netmap adapter to be used by the driver. */
2349 struct netmap_hw_adapter dr;
2351 /* Reference counter to track users of backend netmap port: the
2352 * network stack and netmap clients.
2353 * Used to decide when we need (de)allocate krings/rings and
2354 * start (stop) ptnetmap kthreads. */
2359 int netmap_pt_guest_attach(struct netmap_adapter *na,
2360 unsigned int nifp_offset,
2361 unsigned int memid);
2362 bool netmap_pt_guest_txsync(struct nm_csb_atok *atok,
2363 struct nm_csb_ktoa *ktoa,
2364 struct netmap_kring *kring, int flags);
2365 bool netmap_pt_guest_rxsync(struct nm_csb_atok *atok,
2366 struct nm_csb_ktoa *ktoa,
2367 struct netmap_kring *kring, int flags);
2368 int ptnet_nm_krings_create(struct netmap_adapter *na);
2369 void ptnet_nm_krings_delete(struct netmap_adapter *na);
2370 void ptnet_nm_dtor(struct netmap_adapter *na);
2372 /* Helper function wrapping nm_sync_kloop_appl_read(). */
2374 ptnet_sync_tail(struct nm_csb_ktoa *ktoa, struct netmap_kring *kring)
2376 struct netmap_ring *ring = kring->ring;
2378 /* Update hwcur and hwtail as known by the host. */
2379 nm_sync_kloop_appl_read(ktoa, &kring->nr_hwtail, &kring->nr_hwcur);
2381 /* nm_sync_finalize */
2382 ring->tail = kring->rtail = kring->nr_hwtail;
2384 #endif /* WITH_PTNETMAP */
2388 * FreeBSD mbuf allocator/deallocator in emulation mode:
2390 #if __FreeBSD_version < 1100000
2393 * For older versions of FreeBSD:
2395 * We allocate EXT_PACKET mbuf+clusters, but need to set M_NOFREE
2396 * so that the destructor, if invoked, will not free the packet.
2397 * In principle we should set the destructor only on demand,
2398 * but since there might be a race we better do it on allocation.
2399 * As a consequence, we also need to set the destructor or we
2400 * would leak buffers.
2403 /* mbuf destructor, also need to change the type to EXT_EXTREF,
2404 * add an M_NOFREE flag, and then clear the flag and
2405 * chain into uma_zfree(zone_pack, mf)
2406 * (or reinstall the buffer ?)
2408 #define SET_MBUF_DESTRUCTOR(m, fn) do { \
2409 (m)->m_ext.ext_free = (void *)fn; \
2410 (m)->m_ext.ext_type = EXT_EXTREF; \
2414 void_mbuf_dtor(struct mbuf *m, void *arg1, void *arg2)
2416 /* restore original mbuf */
2417 m->m_ext.ext_buf = m->m_data = m->m_ext.ext_arg1;
2418 m->m_ext.ext_arg1 = NULL;
2419 m->m_ext.ext_type = EXT_PACKET;
2420 m->m_ext.ext_free = NULL;
2421 if (MBUF_REFCNT(m) == 0)
2422 SET_MBUF_REFCNT(m, 1);
2423 uma_zfree(zone_pack, m);
2428 static inline struct mbuf *
2429 nm_os_get_mbuf(struct ifnet *ifp, int len)
2434 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
2436 /* m_getcl() (mb_ctor_mbuf) has an assert that checks that
2437 * M_NOFREE flag is not specified as third argument,
2438 * so we have to set M_NOFREE after m_getcl(). */
2439 m->m_flags |= M_NOFREE;
2440 m->m_ext.ext_arg1 = m->m_ext.ext_buf; // XXX save
2441 m->m_ext.ext_free = (void *)void_mbuf_dtor;
2442 m->m_ext.ext_type = EXT_EXTREF;
2443 nm_prdis(5, "create m %p refcnt %d", m, MBUF_REFCNT(m));
2448 #else /* __FreeBSD_version >= 1100000 */
2451 * Newer versions of FreeBSD, using a straightforward scheme.
2453 * We allocate mbufs with m_gethdr(), since the mbuf header is needed
2454 * by the driver. We also attach a customly-provided external storage,
2455 * which in this case is a netmap buffer. When calling m_extadd(), however
2456 * we pass a NULL address, since the real address (and length) will be
2457 * filled in by nm_os_generic_xmit_frame() right before calling
2460 * The dtor function does nothing, however we need it since mb_free_ext()
2461 * has a KASSERT(), checking that the mbuf dtor function is not NULL.
2464 #if __FreeBSD_version <= 1200050
2465 static void void_mbuf_dtor(struct mbuf *m, void *arg1, void *arg2) { }
2466 #else /* __FreeBSD_version >= 1200051 */
2467 /* The arg1 and arg2 pointers argument were removed by r324446, which
2468 * in included since version 1200051. */
2469 static void void_mbuf_dtor(struct mbuf *m) { }
2470 #endif /* __FreeBSD_version >= 1200051 */
2472 #define SET_MBUF_DESTRUCTOR(m, fn) do { \
2473 (m)->m_ext.ext_free = (fn != NULL) ? \
2474 (void *)fn : (void *)void_mbuf_dtor; \
2477 static inline struct mbuf *
2478 nm_os_get_mbuf(struct ifnet *ifp, int len)
2485 m = m_gethdr(M_NOWAIT, MT_DATA);
2490 m_extadd(m, NULL /* buf */, 0 /* size */, void_mbuf_dtor,
2491 NULL, NULL, 0, EXT_NET_DRV);
2496 #endif /* __FreeBSD_version >= 1100000 */
2497 #endif /* __FreeBSD__ */
2499 struct nmreq_option * nmreq_getoption(struct nmreq_header *, uint16_t);
2501 int netmap_init_bridges(void);
2502 void netmap_uninit_bridges(void);
2504 /* Functions to read and write CSB fields from the kernel. */
2506 #define CSB_READ(csb, field, r) (get_user(r, &csb->field))
2507 #define CSB_WRITE(csb, field, v) (put_user(v, &csb->field))
2509 #define CSB_READ(csb, field, r) (r = fuword32(&csb->field))
2510 #define CSB_WRITE(csb, field, v) (suword32(&csb->field, v))
2511 #endif /* ! linux */
2513 /* some macros that may not be defined */
2518 #define ETH_FCS_LEN 4
2524 #endif /* _NET_NETMAP_KERN_H_ */