xref: /freebsd/sys/dev/netmap/netmap_kern.h (revision 5608fd23c27fa1e8ee595d7b678cbfd35d657fbe)
1 /*
2  * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved.
3  * Copyright (C) 2013-2014 Universita` di Pisa. All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  *   1. Redistributions of source code must retain the above copyright
9  *      notice, this list of conditions and the following disclaimer.
10  *   2. Redistributions in binary form must reproduce the above copyright
11  *      notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 /*
28  * $FreeBSD$
29  *
30  * The header contains the definitions of constants and function
31  * prototypes used only in kernelspace.
32  */
33 
34 #ifndef _NET_NETMAP_KERN_H_
35 #define _NET_NETMAP_KERN_H_
36 
37 #define WITH_VALE	// comment out to disable VALE support
38 #define WITH_PIPES
39 #define WITH_MONITOR
40 
41 #if defined(__FreeBSD__)
42 
43 #define likely(x)	__builtin_expect((long)!!(x), 1L)
44 #define unlikely(x)	__builtin_expect((long)!!(x), 0L)
45 
46 #define	NM_LOCK_T	struct mtx
47 #define	NMG_LOCK_T	struct mtx
48 #define NMG_LOCK_INIT()	mtx_init(&netmap_global_lock, \
49 				"netmap global lock", NULL, MTX_DEF)
50 #define NMG_LOCK_DESTROY()	mtx_destroy(&netmap_global_lock)
51 #define NMG_LOCK()	mtx_lock(&netmap_global_lock)
52 #define NMG_UNLOCK()	mtx_unlock(&netmap_global_lock)
53 #define NMG_LOCK_ASSERT()	mtx_assert(&netmap_global_lock, MA_OWNED)
54 
55 #define	NM_SELINFO_T	struct selinfo
56 #define	MBUF_LEN(m)	((m)->m_pkthdr.len)
57 #define	MBUF_IFP(m)	((m)->m_pkthdr.rcvif)
58 #define	NM_SEND_UP(ifp, m)	((NA(ifp))->if_input)(ifp, m)
59 
60 #define NM_ATOMIC_T	volatile int	// XXX ?
61 /* atomic operations */
62 #include <machine/atomic.h>
63 #define NM_ATOMIC_TEST_AND_SET(p)       (!atomic_cmpset_acq_int((p), 0, 1))
64 #define NM_ATOMIC_CLEAR(p)              atomic_store_rel_int((p), 0)
65 
66 #if __FreeBSD_version >= 1100005
67 struct netmap_adapter *netmap_getna(if_t ifp);
68 #endif
69 
70 #if __FreeBSD_version >= 1100027
71 #define GET_MBUF_REFCNT(m)      ((m)->m_ext.ext_cnt ? *((m)->m_ext.ext_cnt) : -1)
72 #define SET_MBUF_REFCNT(m, x)   *((m)->m_ext.ext_cnt) = x
73 #define PNT_MBUF_REFCNT(m)      ((m)->m_ext.ext_cnt)
74 #else
75 #define GET_MBUF_REFCNT(m)      ((m)->m_ext.ref_cnt ? *((m)->m_ext.ref_cnt) : -1)
76 #define SET_MBUF_REFCNT(m, x)   *((m)->m_ext.ref_cnt) = x
77 #define PNT_MBUF_REFCNT(m)      ((m)->m_ext.ref_cnt)
78 #endif
79 
80 MALLOC_DECLARE(M_NETMAP);
81 
82 // XXX linux struct, not used in FreeBSD
83 struct net_device_ops {
84 };
85 struct ethtool_ops {
86 };
87 struct hrtimer {
88 };
89 
90 #elif defined (linux)
91 
92 #define	NM_LOCK_T	safe_spinlock_t	// see bsd_glue.h
93 #define	NM_SELINFO_T	wait_queue_head_t
94 #define	MBUF_LEN(m)	((m)->len)
95 #define	MBUF_IFP(m)	((m)->dev)
96 #define	NM_SEND_UP(ifp, m)  \
97                         do { \
98                             m->priority = NM_MAGIC_PRIORITY_RX; \
99                             netif_rx(m); \
100                         } while (0)
101 
102 #define NM_ATOMIC_T	volatile long unsigned int
103 
104 // XXX a mtx would suffice here too 20130404 gl
105 #define NMG_LOCK_T		struct semaphore
106 #define NMG_LOCK_INIT()		sema_init(&netmap_global_lock, 1)
107 #define NMG_LOCK_DESTROY()
108 #define NMG_LOCK()		down(&netmap_global_lock)
109 #define NMG_UNLOCK()		up(&netmap_global_lock)
110 #define NMG_LOCK_ASSERT()	//	XXX to be completed
111 
112 #ifndef DEV_NETMAP
113 #define DEV_NETMAP
114 #endif /* DEV_NETMAP */
115 
116 #elif defined (__APPLE__)
117 
118 #warning apple support is incomplete.
119 #define likely(x)	__builtin_expect(!!(x), 1)
120 #define unlikely(x)	__builtin_expect(!!(x), 0)
121 #define	NM_LOCK_T	IOLock *
122 #define	NM_SELINFO_T	struct selinfo
123 #define	MBUF_LEN(m)	((m)->m_pkthdr.len)
124 #define	NM_SEND_UP(ifp, m)	((ifp)->if_input)(ifp, m)
125 
126 #else
127 
128 #error unsupported platform
129 
130 #endif /* end - platform-specific code */
131 
132 #define ND(format, ...)
133 #define D(format, ...)						\
134 	do {							\
135 		struct timeval __xxts;				\
136 		microtime(&__xxts);				\
137 		printf("%03d.%06d [%4d] %-25s " format "\n",	\
138 		(int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec,	\
139 		__LINE__, __FUNCTION__, ##__VA_ARGS__);		\
140 	} while (0)
141 
142 /* rate limited, lps indicates how many per second */
143 #define RD(lps, format, ...)					\
144 	do {							\
145 		static int t0, __cnt;				\
146 		if (t0 != time_second) {			\
147 			t0 = time_second;			\
148 			__cnt = 0;				\
149 		}						\
150 		if (__cnt++ < lps)				\
151 			D(format, ##__VA_ARGS__);		\
152 	} while (0)
153 
154 struct netmap_adapter;
155 struct nm_bdg_fwd;
156 struct nm_bridge;
157 struct netmap_priv_d;
158 
159 const char *nm_dump_buf(char *p, int len, int lim, char *dst);
160 
161 #include "netmap_mbq.h"
162 
163 extern NMG_LOCK_T	netmap_global_lock;
164 
165 /*
166  * private, kernel view of a ring. Keeps track of the status of
167  * a ring across system calls.
168  *
169  *	nr_hwcur	index of the next buffer to refill.
170  *			It corresponds to ring->head
171  *			at the time the system call returns.
172  *
173  *	nr_hwtail	index of the first buffer owned by the kernel.
174  *			On RX, hwcur->hwtail are receive buffers
175  *			not yet released. hwcur is advanced following
176  *			ring->head, hwtail is advanced on incoming packets,
177  *			and a wakeup is generated when hwtail passes ring->cur
178  *			    On TX, hwcur->rcur have been filled by the sender
179  *			but not sent yet to the NIC; rcur->hwtail are available
180  *			for new transmissions, and hwtail->hwcur-1 are pending
181  *			transmissions not yet acknowledged.
182  *
183  * The indexes in the NIC and netmap rings are offset by nkr_hwofs slots.
184  * This is so that, on a reset, buffers owned by userspace are not
185  * modified by the kernel. In particular:
186  * RX rings: the next empty buffer (hwtail + hwofs) coincides with
187  * 	the next empty buffer as known by the hardware (next_to_check or so).
188  * TX rings: hwcur + hwofs coincides with next_to_send
189  *
190  * For received packets, slot->flags is set to nkr_slot_flags
191  * so we can provide a proper initial value (e.g. set NS_FORWARD
192  * when operating in 'transparent' mode).
193  *
194  * The following fields are used to implement lock-free copy of packets
195  * from input to output ports in VALE switch:
196  *	nkr_hwlease	buffer after the last one being copied.
197  *			A writer in nm_bdg_flush reserves N buffers
198  *			from nr_hwlease, advances it, then does the
199  *			copy outside the lock.
200  *			In RX rings (used for VALE ports),
201  *			nkr_hwtail <= nkr_hwlease < nkr_hwcur+N-1
202  *			In TX rings (used for NIC or host stack ports)
203  *			nkr_hwcur <= nkr_hwlease < nkr_hwtail
204  *	nkr_leases	array of nkr_num_slots where writers can report
205  *			completion of their block. NR_NOSLOT (~0) indicates
206  *			that the writer has not finished yet
207  *	nkr_lease_idx	index of next free slot in nr_leases, to be assigned
208  *
209  * The kring is manipulated by txsync/rxsync and generic netmap function.
210  *
211  * Concurrent rxsync or txsync on the same ring are prevented through
212  * by nm_kr_(try)lock() which in turn uses nr_busy. This is all we need
213  * for NIC rings, and for TX rings attached to the host stack.
214  *
215  * RX rings attached to the host stack use an mbq (rx_queue) on both
216  * rxsync_from_host() and netmap_transmit(). The mbq is protected
217  * by its internal lock.
218  *
219  * RX rings attached to the VALE switch are accessed by both senders
220  * and receiver. They are protected through the q_lock on the RX ring.
221  */
222 struct netmap_kring {
223 	struct netmap_ring	*ring;
224 
225 	uint32_t	nr_hwcur;
226 	uint32_t	nr_hwtail;
227 
228 	/*
229 	 * Copies of values in user rings, so we do not need to look
230 	 * at the ring (which could be modified). These are set in the
231 	 * *sync_prologue()/finalize() routines.
232 	 */
233 	uint32_t	rhead;
234 	uint32_t	rcur;
235 	uint32_t	rtail;
236 
237 	uint32_t	nr_kflags;	/* private driver flags */
238 #define NKR_PENDINTR	0x1		// Pending interrupt.
239 	uint32_t	nkr_num_slots;
240 
241 	/*
242 	 * On a NIC reset, the NIC ring indexes may be reset but the
243 	 * indexes in the netmap rings remain the same. nkr_hwofs
244 	 * keeps track of the offset between the two.
245 	 */
246 	int32_t		nkr_hwofs;
247 
248 	uint16_t	nkr_slot_flags;	/* initial value for flags */
249 
250 	/* last_reclaim is opaque marker to help reduce the frequency
251 	 * of operations such as reclaiming tx buffers. A possible use
252 	 * is set it to ticks and do the reclaim only once per tick.
253 	 */
254 	uint64_t	last_reclaim;
255 
256 
257 	NM_SELINFO_T	si;		/* poll/select wait queue */
258 	NM_LOCK_T	q_lock;		/* protects kring and ring. */
259 	NM_ATOMIC_T	nr_busy;	/* prevent concurrent syscalls */
260 
261 	struct netmap_adapter *na;
262 
263 	/* The folloiwing fields are for VALE switch support */
264 	struct nm_bdg_fwd *nkr_ft;
265 	uint32_t	*nkr_leases;
266 #define NR_NOSLOT	((uint32_t)~0)	/* used in nkr_*lease* */
267 	uint32_t	nkr_hwlease;
268 	uint32_t	nkr_lease_idx;
269 
270 	/* while nkr_stopped is set, no new [tr]xsync operations can
271 	 * be started on this kring.
272 	 * This is used by netmap_disable_all_rings()
273 	 * to find a synchronization point where critical data
274 	 * structures pointed to by the kring can be added or removed
275 	 */
276 	volatile int nkr_stopped;
277 
278 	/* Support for adapters without native netmap support.
279 	 * On tx rings we preallocate an array of tx buffers
280 	 * (same size as the netmap ring), on rx rings we
281 	 * store incoming mbufs in a queue that is drained by
282 	 * a rxsync.
283 	 */
284 	struct mbuf **tx_pool;
285 	// u_int nr_ntc;		/* Emulation of a next-to-clean RX ring pointer. */
286 	struct mbq rx_queue;            /* intercepted rx mbufs. */
287 
288 	uint32_t	ring_id;	/* debugging */
289 	char name[64];			/* diagnostic */
290 
291 	/* [tx]sync callback for this kring.
292 	 * The default nm_kring_create callback (netmap_krings_create)
293 	 * sets the nm_sync callback of each hardware tx(rx) kring to
294 	 * the corresponding nm_txsync(nm_rxsync) taken from the
295 	 * netmap_adapter; moreover, it sets the sync callback
296 	 * of the host tx(rx) ring to netmap_txsync_to_host
297 	 * (netmap_rxsync_from_host).
298 	 *
299 	 * Overrides: the above configuration is not changed by
300 	 * any of the nm_krings_create callbacks.
301 	 */
302 	int (*nm_sync)(struct netmap_kring *kring, int flags);
303 
304 #ifdef WITH_PIPES
305 	struct netmap_kring *pipe;	/* if this is a pipe ring,
306 					 * pointer to the other end
307 					 */
308 	struct netmap_ring *save_ring;	/* pointer to hidden rings
309        					 * (see netmap_pipe.c for details)
310 					 */
311 #endif /* WITH_PIPES */
312 
313 #ifdef WITH_MONITOR
314 	/* pointer to the adapter that is monitoring this kring (if any)
315 	 */
316 	struct netmap_monitor_adapter *monitor;
317 	/*
318 	 * Monitors work by intercepting the txsync and/or rxsync of the
319 	 * monitored krings. This is implemented by replacing
320 	 * the nm_sync pointer above and saving the previous
321 	 * one in save_sync below.
322 	 */
323 	int (*save_sync)(struct netmap_kring *kring, int flags);
324 #endif
325 } __attribute__((__aligned__(64)));
326 
327 
328 /* return the next index, with wraparound */
329 static inline uint32_t
330 nm_next(uint32_t i, uint32_t lim)
331 {
332 	return unlikely (i == lim) ? 0 : i + 1;
333 }
334 
335 
336 /* return the previous index, with wraparound */
337 static inline uint32_t
338 nm_prev(uint32_t i, uint32_t lim)
339 {
340 	return unlikely (i == 0) ? lim : i - 1;
341 }
342 
343 
344 /*
345  *
346  * Here is the layout for the Rx and Tx rings.
347 
348        RxRING                            TxRING
349 
350       +-----------------+            +-----------------+
351       |                 |            |                 |
352       |XXX free slot XXX|            |XXX free slot XXX|
353       +-----------------+            +-----------------+
354 head->| owned by user   |<-hwcur     | not sent to nic |<-hwcur
355       |                 |            | yet             |
356       +-----------------+            |                 |
357  cur->| available to    |            |                 |
358       | user, not read  |            +-----------------+
359       | yet             |       cur->| (being          |
360       |                 |            |  prepared)      |
361       |                 |            |                 |
362       +-----------------+            +     ------      +
363 tail->|                 |<-hwtail    |                 |<-hwlease
364       | (being          | ...        |                 | ...
365       |  prepared)      | ...        |                 | ...
366       +-----------------+ ...        |                 | ...
367       |                 |<-hwlease   +-----------------+
368       |                 |      tail->|                 |<-hwtail
369       |                 |            |                 |
370       |                 |            |                 |
371       |                 |            |                 |
372       +-----------------+            +-----------------+
373 
374  * The cur/tail (user view) and hwcur/hwtail (kernel view)
375  * are used in the normal operation of the card.
376  *
377  * When a ring is the output of a switch port (Rx ring for
378  * a VALE port, Tx ring for the host stack or NIC), slots
379  * are reserved in blocks through 'hwlease' which points
380  * to the next unused slot.
381  * On an Rx ring, hwlease is always after hwtail,
382  * and completions cause hwtail to advance.
383  * On a Tx ring, hwlease is always between cur and hwtail,
384  * and completions cause cur to advance.
385  *
386  * nm_kr_space() returns the maximum number of slots that
387  * can be assigned.
388  * nm_kr_lease() reserves the required number of buffers,
389  *    advances nkr_hwlease and also returns an entry in
390  *    a circular array where completions should be reported.
391  */
392 
393 
394 
395 enum txrx { NR_RX = 0, NR_TX = 1 };
396 
397 struct netmap_vp_adapter; // forward
398 
399 /*
400  * The "struct netmap_adapter" extends the "struct adapter"
401  * (or equivalent) device descriptor.
402  * It contains all base fields needed to support netmap operation.
403  * There are in fact different types of netmap adapters
404  * (native, generic, VALE switch...) so a netmap_adapter is
405  * just the first field in the derived type.
406  */
407 struct netmap_adapter {
408 	/*
409 	 * On linux we do not have a good way to tell if an interface
410 	 * is netmap-capable. So we always use the following trick:
411 	 * NA(ifp) points here, and the first entry (which hopefully
412 	 * always exists and is at least 32 bits) contains a magic
413 	 * value which we can use to detect that the interface is good.
414 	 */
415 	uint32_t magic;
416 	uint32_t na_flags;	/* enabled, and other flags */
417 #define NAF_SKIP_INTR	1	/* use the regular interrupt handler.
418 				 * useful during initialization
419 				 */
420 #define NAF_SW_ONLY	2	/* forward packets only to sw adapter */
421 #define NAF_BDG_MAYSLEEP 4	/* the bridge is allowed to sleep when
422 				 * forwarding packets coming from this
423 				 * interface
424 				 */
425 #define NAF_MEM_OWNER	8	/* the adapter is responsible for the
426 				 * deallocation of the memory allocator
427 				 */
428 #define NAF_NATIVE_ON   16      /* the adapter is native and the attached
429 				 * interface is in netmap mode.
430 				 * Virtual ports (vale, pipe, monitor...)
431 				 * should never use this flag.
432 				 */
433 #define	NAF_NETMAP_ON	32	/* netmap is active (either native or
434 				 * emulated). Where possible (e.g. FreeBSD)
435 				 * IFCAP_NETMAP also mirrors this flag.
436 				 */
437 #define NAF_HOST_RINGS  64	/* the adapter supports the host rings */
438 #define NAF_FORCE_NATIVE 128	/* the adapter is always NATIVE */
439 #define	NAF_BUSY	(1U<<31) /* the adapter is used internally and
440 				  * cannot be registered from userspace
441 				  */
442 	int active_fds; /* number of user-space descriptors using this
443 			 interface, which is equal to the number of
444 			 struct netmap_if objs in the mapped region. */
445 
446 	u_int num_rx_rings; /* number of adapter receive rings */
447 	u_int num_tx_rings; /* number of adapter transmit rings */
448 
449 	u_int num_tx_desc; /* number of descriptor in each queue */
450 	u_int num_rx_desc;
451 
452 	/* tx_rings and rx_rings are private but allocated
453 	 * as a contiguous chunk of memory. Each array has
454 	 * N+1 entries, for the adapter queues and for the host queue.
455 	 */
456 	struct netmap_kring *tx_rings; /* array of TX rings. */
457 	struct netmap_kring *rx_rings; /* array of RX rings. */
458 
459 	void *tailroom;		       /* space below the rings array */
460 				       /* (used for leases) */
461 
462 
463 	NM_SELINFO_T tx_si, rx_si;	/* global wait queues */
464 
465 	/* count users of the global wait queues */
466 	int tx_si_users, rx_si_users;
467 
468 	void *pdev; /* used to store pci device */
469 
470 	/* copy of if_qflush and if_transmit pointers, to intercept
471 	 * packets from the network stack when netmap is active.
472 	 */
473 	int     (*if_transmit)(struct ifnet *, struct mbuf *);
474 
475 	/* copy of if_input for netmap_send_up() */
476 	void     (*if_input)(struct ifnet *, struct mbuf *);
477 
478 	/* references to the ifnet and device routines, used by
479 	 * the generic netmap functions.
480 	 */
481 	struct ifnet *ifp; /* adapter is ifp->if_softc */
482 
483 	/*---- callbacks for this netmap adapter -----*/
484 	/*
485 	 * nm_dtor() is the cleanup routine called when destroying
486 	 *	the adapter.
487 	 *	Called with NMG_LOCK held.
488 	 *
489 	 * nm_register() is called on NIOCREGIF and close() to enter
490 	 *	or exit netmap mode on the NIC
491 	 *	Called with NNG_LOCK held.
492 	 *
493 	 * nm_txsync() pushes packets to the underlying hw/switch
494 	 *
495 	 * nm_rxsync() collects packets from the underlying hw/switch
496 	 *
497 	 * nm_config() returns configuration information from the OS
498 	 *	Called with NMG_LOCK held.
499 	 *
500 	 * nm_krings_create() create and init the tx_rings and
501 	 * 	rx_rings arrays of kring structures. In particular,
502 	 * 	set the nm_sync callbacks for each ring.
503 	 * 	There is no need to also allocate the corresponding
504 	 * 	netmap_rings, since netmap_mem_rings_create() will always
505 	 * 	be called to provide the missing ones.
506 	 *	Called with NNG_LOCK held.
507 	 *
508 	 * nm_krings_delete() cleanup and delete the tx_rings and rx_rings
509 	 * 	arrays
510 	 *	Called with NMG_LOCK held.
511 	 *
512 	 * nm_notify() is used to act after data have become available
513 	 * 	(or the stopped state of the ring has changed)
514 	 *	For hw devices this is typically a selwakeup(),
515 	 *	but for NIC/host ports attached to a switch (or vice-versa)
516 	 *	we also need to invoke the 'txsync' code downstream.
517 	 */
518 	void (*nm_dtor)(struct netmap_adapter *);
519 
520 	int (*nm_register)(struct netmap_adapter *, int onoff);
521 
522 	int (*nm_txsync)(struct netmap_kring *kring, int flags);
523 	int (*nm_rxsync)(struct netmap_kring *kring, int flags);
524 #define NAF_FORCE_READ    1
525 #define NAF_FORCE_RECLAIM 2
526 	/* return configuration information */
527 	int (*nm_config)(struct netmap_adapter *,
528 		u_int *txr, u_int *txd, u_int *rxr, u_int *rxd);
529 	int (*nm_krings_create)(struct netmap_adapter *);
530 	void (*nm_krings_delete)(struct netmap_adapter *);
531 	int (*nm_notify)(struct netmap_adapter *,
532 		u_int ring, enum txrx, int flags);
533 #define NAF_DISABLE_NOTIFY 8	/* notify that the stopped state of the
534 				 * ring has changed (kring->nkr_stopped)
535 				 */
536 
537 #ifdef WITH_VALE
538 	/*
539 	 * nm_bdg_attach() initializes the na_vp field to point
540 	 *      to an adapter that can be attached to a VALE switch. If the
541 	 *      current adapter is already a VALE port, na_vp is simply a cast;
542 	 *      otherwise, na_vp points to a netmap_bwrap_adapter.
543 	 *      If applicable, this callback also initializes na_hostvp,
544 	 *      that can be used to connect the adapter host rings to the
545 	 *      switch.
546 	 *      Called with NMG_LOCK held.
547 	 *
548 	 * nm_bdg_ctl() is called on the actual attach/detach to/from
549 	 *      to/from the switch, to perform adapter-specific
550 	 *      initializations
551 	 *      Called with NMG_LOCK held.
552 	 */
553 	int (*nm_bdg_attach)(const char *bdg_name, struct netmap_adapter *);
554 	int (*nm_bdg_ctl)(struct netmap_adapter *, struct nmreq *, int);
555 
556 	/* adapter used to attach this adapter to a VALE switch (if any) */
557 	struct netmap_vp_adapter *na_vp;
558 	/* adapter used to attach the host rings of this adapter
559 	 * to a VALE switch (if any) */
560 	struct netmap_vp_adapter *na_hostvp;
561 #endif
562 
563 	/* standard refcount to control the lifetime of the adapter
564 	 * (it should be equal to the lifetime of the corresponding ifp)
565 	 */
566 	int na_refcount;
567 
568 	/* memory allocator (opaque)
569 	 * We also cache a pointer to the lut_entry for translating
570 	 * buffer addresses, and the total number of buffers.
571 	 */
572  	struct netmap_mem_d *nm_mem;
573 	struct lut_entry *na_lut;
574 	uint32_t na_lut_objtotal;	/* max buffer index */
575 	uint32_t na_lut_objsize;	/* buffer size */
576 
577 	/* additional information attached to this adapter
578 	 * by other netmap subsystems. Currently used by
579 	 * bwrap and LINUX/v1000.
580 	 */
581 	void *na_private;
582 
583 #ifdef WITH_PIPES
584 	/* array of pipes that have this adapter as a parent */
585 	struct netmap_pipe_adapter **na_pipes;
586 	int na_next_pipe;	/* next free slot in the array */
587 	int na_max_pipes;	/* size of the array */
588 #endif /* WITH_PIPES */
589 
590 	char name[64];
591 };
592 
593 
594 /*
595  * If the NIC is owned by the kernel
596  * (i.e., bridge), neither another bridge nor user can use it;
597  * if the NIC is owned by a user, only users can share it.
598  * Evaluation must be done under NMG_LOCK().
599  */
600 #define NETMAP_OWNED_BY_KERN(na)	((na)->na_flags & NAF_BUSY)
601 #define NETMAP_OWNED_BY_ANY(na) \
602 	(NETMAP_OWNED_BY_KERN(na) || ((na)->active_fds > 0))
603 
604 
605 /*
606  * derived netmap adapters for various types of ports
607  */
608 struct netmap_vp_adapter {	/* VALE software port */
609 	struct netmap_adapter up;
610 
611 	/*
612 	 * Bridge support:
613 	 *
614 	 * bdg_port is the port number used in the bridge;
615 	 * na_bdg points to the bridge this NA is attached to.
616 	 */
617 	int bdg_port;
618 	struct nm_bridge *na_bdg;
619 	int retry;
620 
621 	/* Offset of ethernet header for each packet. */
622 	u_int virt_hdr_len;
623 	/* Maximum Frame Size, used in bdg_mismatch_datapath() */
624 	u_int mfs;
625 };
626 
627 
628 struct netmap_hw_adapter {	/* physical device */
629 	struct netmap_adapter up;
630 
631 	struct net_device_ops nm_ndo;	// XXX linux only
632 	struct ethtool_ops    nm_eto;	// XXX linux only
633 	const struct ethtool_ops*   save_ethtool;
634 
635 	int (*nm_hw_register)(struct netmap_adapter *, int onoff);
636 };
637 
638 /* Mitigation support. */
639 struct nm_generic_mit {
640 	struct hrtimer mit_timer;
641 	int mit_pending;
642 	int mit_ring_idx;  /* index of the ring being mitigated */
643 	struct netmap_adapter *mit_na;  /* backpointer */
644 };
645 
646 struct netmap_generic_adapter {	/* emulated device */
647 	struct netmap_hw_adapter up;
648 
649 	/* Pointer to a previously used netmap adapter. */
650 	struct netmap_adapter *prev;
651 
652 	/* generic netmap adapters support:
653 	 * a net_device_ops struct overrides ndo_select_queue(),
654 	 * save_if_input saves the if_input hook (FreeBSD),
655 	 * mit implements rx interrupt mitigation,
656 	 */
657 	struct net_device_ops generic_ndo;
658 	void (*save_if_input)(struct ifnet *, struct mbuf *);
659 
660 	struct nm_generic_mit *mit;
661 #ifdef linux
662         netdev_tx_t (*save_start_xmit)(struct mbuf *, struct ifnet *);
663 #endif
664 };
665 
666 static __inline int
667 netmap_real_tx_rings(struct netmap_adapter *na)
668 {
669 	return na->num_tx_rings + !!(na->na_flags & NAF_HOST_RINGS);
670 }
671 
672 static __inline int
673 netmap_real_rx_rings(struct netmap_adapter *na)
674 {
675 	return na->num_rx_rings + !!(na->na_flags & NAF_HOST_RINGS);
676 }
677 
678 #ifdef WITH_VALE
679 
680 /*
681  * Bridge wrapper for non VALE ports attached to a VALE switch.
682  *
683  * The real device must already have its own netmap adapter (hwna).
684  * The bridge wrapper and the hwna adapter share the same set of
685  * netmap rings and buffers, but they have two separate sets of
686  * krings descriptors, with tx/rx meanings swapped:
687  *
688  *                                  netmap
689  *           bwrap     krings       rings      krings      hwna
690  *         +------+   +------+     +-----+    +------+   +------+
691  *         |tx_rings->|      |\   /|     |----|      |<-tx_rings|
692  *         |      |   +------+ \ / +-----+    +------+   |      |
693  *         |      |             X                        |      |
694  *         |      |            / \                       |      |
695  *         |      |   +------+/   \+-----+    +------+   |      |
696  *         |rx_rings->|      |     |     |----|      |<-rx_rings|
697  *         |      |   +------+     +-----+    +------+   |      |
698  *         +------+                                      +------+
699  *
700  * - packets coming from the bridge go to the brwap rx rings,
701  *   which are also the hwna tx rings.  The bwrap notify callback
702  *   will then complete the hwna tx (see netmap_bwrap_notify).
703  *
704  * - packets coming from the outside go to the hwna rx rings,
705  *   which are also the bwrap tx rings.  The (overwritten) hwna
706  *   notify method will then complete the bridge tx
707  *   (see netmap_bwrap_intr_notify).
708  *
709  *   The bridge wrapper may optionally connect the hwna 'host' rings
710  *   to the bridge. This is done by using a second port in the
711  *   bridge and connecting it to the 'host' netmap_vp_adapter
712  *   contained in the netmap_bwrap_adapter. The brwap host adapter
713  *   cross-links the hwna host rings in the same way as shown above.
714  *
715  * - packets coming from the bridge and directed to the host stack
716  *   are handled by the bwrap host notify callback
717  *   (see netmap_bwrap_host_notify)
718  *
719  * - packets coming from the host stack are still handled by the
720  *   overwritten hwna notify callback (netmap_bwrap_intr_notify),
721  *   but are diverted to the host adapter depending on the ring number.
722  *
723  */
724 struct netmap_bwrap_adapter {
725 	struct netmap_vp_adapter up;
726 	struct netmap_vp_adapter host;  /* for host rings */
727 	struct netmap_adapter *hwna;	/* the underlying device */
728 
729 	/* backup of the hwna notify callback */
730 	int (*save_notify)(struct netmap_adapter *,
731 			u_int ring, enum txrx, int flags);
732 	/* backup of the hwna memory allocator */
733 	struct netmap_mem_d *save_nmd;
734 
735 	/*
736 	 * When we attach a physical interface to the bridge, we
737 	 * allow the controlling process to terminate, so we need
738 	 * a place to store the n_detmap_priv_d data structure.
739 	 * This is only done when physical interfaces
740 	 * are attached to a bridge.
741 	 */
742 	struct netmap_priv_d *na_kpriv;
743 };
744 int netmap_bwrap_attach(const char *name, struct netmap_adapter *);
745 
746 
747 #endif /* WITH_VALE */
748 
749 #ifdef WITH_PIPES
750 
751 #define NM_MAXPIPES 	64	/* max number of pipes per adapter */
752 
753 struct netmap_pipe_adapter {
754 	struct netmap_adapter up;
755 
756 	u_int id; 	/* pipe identifier */
757 	int role;	/* either NR_REG_PIPE_MASTER or NR_REG_PIPE_SLAVE */
758 
759 	struct netmap_adapter *parent; /* adapter that owns the memory */
760 	struct netmap_pipe_adapter *peer; /* the other end of the pipe */
761 	int peer_ref;		/* 1 iff we are holding a ref to the peer */
762 
763 	u_int parent_slot; /* index in the parent pipe array */
764 };
765 
766 #endif /* WITH_PIPES */
767 
768 
769 /* return slots reserved to rx clients; used in drivers */
770 static inline uint32_t
771 nm_kr_rxspace(struct netmap_kring *k)
772 {
773 	int space = k->nr_hwtail - k->nr_hwcur;
774 	if (space < 0)
775 		space += k->nkr_num_slots;
776 	ND("preserving %d rx slots %d -> %d", space, k->nr_hwcur, k->nr_hwtail);
777 
778 	return space;
779 }
780 
781 
782 /* True if no space in the tx ring. only valid after txsync_prologue */
783 static inline int
784 nm_kr_txempty(struct netmap_kring *kring)
785 {
786 	return kring->rcur == kring->nr_hwtail;
787 }
788 
789 
790 /*
791  * protect against multiple threads using the same ring.
792  * also check that the ring has not been stopped.
793  * We only care for 0 or !=0 as a return code.
794  */
795 #define NM_KR_BUSY	1
796 #define NM_KR_STOPPED	2
797 
798 
799 static __inline void nm_kr_put(struct netmap_kring *kr)
800 {
801 	NM_ATOMIC_CLEAR(&kr->nr_busy);
802 }
803 
804 
805 static __inline int nm_kr_tryget(struct netmap_kring *kr)
806 {
807 	/* check a first time without taking the lock
808 	 * to avoid starvation for nm_kr_get()
809 	 */
810 	if (unlikely(kr->nkr_stopped)) {
811 		ND("ring %p stopped (%d)", kr, kr->nkr_stopped);
812 		return NM_KR_STOPPED;
813 	}
814 	if (unlikely(NM_ATOMIC_TEST_AND_SET(&kr->nr_busy)))
815 		return NM_KR_BUSY;
816 	/* check a second time with lock held */
817 	if (unlikely(kr->nkr_stopped)) {
818 		ND("ring %p stopped (%d)", kr, kr->nkr_stopped);
819 		nm_kr_put(kr);
820 		return NM_KR_STOPPED;
821 	}
822 	return 0;
823 }
824 
825 
826 /*
827  * The following functions are used by individual drivers to
828  * support netmap operation.
829  *
830  * netmap_attach() initializes a struct netmap_adapter, allocating the
831  * 	struct netmap_ring's and the struct selinfo.
832  *
833  * netmap_detach() frees the memory allocated by netmap_attach().
834  *
835  * netmap_transmit() replaces the if_transmit routine of the interface,
836  *	and is used to intercept packets coming from the stack.
837  *
838  * netmap_load_map/netmap_reload_map are helper routines to set/reset
839  *	the dmamap for a packet buffer
840  *
841  * netmap_reset() is a helper routine to be called in the hw driver
842  *	when reinitializing a ring. It should not be called by
843  *	virtual ports (vale, pipes, monitor)
844  */
845 int netmap_attach(struct netmap_adapter *);
846 void netmap_detach(struct ifnet *);
847 int netmap_transmit(struct ifnet *, struct mbuf *);
848 struct netmap_slot *netmap_reset(struct netmap_adapter *na,
849 	enum txrx tx, u_int n, u_int new_cur);
850 int netmap_ring_reinit(struct netmap_kring *);
851 
852 /* default functions to handle rx/tx interrupts */
853 int netmap_rx_irq(struct ifnet *, u_int, u_int *);
854 #define netmap_tx_irq(_n, _q) netmap_rx_irq(_n, _q, NULL)
855 void netmap_common_irq(struct ifnet *, u_int, u_int *work_done);
856 
857 
858 #ifdef WITH_VALE
859 /* functions used by external modules to interface with VALE */
860 #define netmap_vp_to_ifp(_vp)	((_vp)->up.ifp)
861 #define netmap_ifp_to_vp(_ifp)	(NA(_ifp)->na_vp)
862 #define netmap_ifp_to_host_vp(_ifp) (NA(_ifp)->na_hostvp)
863 #define netmap_bdg_idx(_vp)	((_vp)->bdg_port)
864 const char *netmap_bdg_name(struct netmap_vp_adapter *);
865 #else /* !WITH_VALE */
866 #define netmap_vp_to_ifp(_vp)	NULL
867 #define netmap_ifp_to_vp(_ifp)	NULL
868 #define netmap_ifp_to_host_vp(_ifp) NULL
869 #define netmap_bdg_idx(_vp)	-1
870 #define netmap_bdg_name(_vp)	NULL
871 #endif /* WITH_VALE */
872 
873 static inline int
874 nm_native_on(struct netmap_adapter *na)
875 {
876 	return na && na->na_flags & NAF_NATIVE_ON;
877 }
878 
879 static inline int
880 nm_netmap_on(struct netmap_adapter *na)
881 {
882 	return na && na->na_flags & NAF_NETMAP_ON;
883 }
884 
885 /* set/clear native flags and if_transmit/netdev_ops */
886 static inline void
887 nm_set_native_flags(struct netmap_adapter *na)
888 {
889 	struct ifnet *ifp = na->ifp;
890 
891 	na->na_flags |= (NAF_NATIVE_ON | NAF_NETMAP_ON);
892 #ifdef IFCAP_NETMAP /* or FreeBSD ? */
893 	ifp->if_capenable |= IFCAP_NETMAP;
894 #endif
895 #ifdef __FreeBSD__
896 	na->if_transmit = ifp->if_transmit;
897 	ifp->if_transmit = netmap_transmit;
898 #else
899 	na->if_transmit = (void *)ifp->netdev_ops;
900 	ifp->netdev_ops = &((struct netmap_hw_adapter *)na)->nm_ndo;
901 	((struct netmap_hw_adapter *)na)->save_ethtool = ifp->ethtool_ops;
902 	ifp->ethtool_ops = &((struct netmap_hw_adapter*)na)->nm_eto;
903 #endif
904 }
905 
906 
907 static inline void
908 nm_clear_native_flags(struct netmap_adapter *na)
909 {
910 	struct ifnet *ifp = na->ifp;
911 
912 #ifdef __FreeBSD__
913 	ifp->if_transmit = na->if_transmit;
914 #else
915 	ifp->netdev_ops = (void *)na->if_transmit;
916 	ifp->ethtool_ops = ((struct netmap_hw_adapter*)na)->save_ethtool;
917 #endif
918 	na->na_flags &= ~(NAF_NATIVE_ON | NAF_NETMAP_ON);
919 #ifdef IFCAP_NETMAP /* or FreeBSD ? */
920 	ifp->if_capenable &= ~IFCAP_NETMAP;
921 #endif
922 }
923 
924 
925 /*
926  * validates parameters in the ring/kring, returns a value for head
927  * If any error, returns ring_size to force a reinit.
928  */
929 uint32_t nm_txsync_prologue(struct netmap_kring *);
930 
931 
932 /*
933  * validates parameters in the ring/kring, returns a value for head,
934  * and the 'reserved' value in the argument.
935  * If any error, returns ring_size lim to force a reinit.
936  */
937 uint32_t nm_rxsync_prologue(struct netmap_kring *);
938 
939 
940 /*
941  * update kring and ring at the end of txsync.
942  */
943 static inline void
944 nm_txsync_finalize(struct netmap_kring *kring)
945 {
946 	/* update ring tail to what the kernel knows */
947 	kring->ring->tail = kring->rtail = kring->nr_hwtail;
948 
949 	/* note, head/rhead/hwcur might be behind cur/rcur
950 	 * if no carrier
951 	 */
952 	ND(5, "%s now hwcur %d hwtail %d head %d cur %d tail %d",
953 		kring->name, kring->nr_hwcur, kring->nr_hwtail,
954 		kring->rhead, kring->rcur, kring->rtail);
955 }
956 
957 
958 /*
959  * update kring and ring at the end of rxsync
960  */
961 static inline void
962 nm_rxsync_finalize(struct netmap_kring *kring)
963 {
964 	/* tell userspace that there might be new packets */
965 	//struct netmap_ring *ring = kring->ring;
966 	ND("head %d cur %d tail %d -> %d", ring->head, ring->cur, ring->tail,
967 		kring->nr_hwtail);
968 	kring->ring->tail = kring->rtail = kring->nr_hwtail;
969 	/* make a copy of the state for next round */
970 	kring->rhead = kring->ring->head;
971 	kring->rcur = kring->ring->cur;
972 }
973 
974 
975 /* check/fix address and len in tx rings */
976 #if 1 /* debug version */
977 #define	NM_CHECK_ADDR_LEN(_na, _a, _l)	do {				\
978 	if (_a == NETMAP_BUF_BASE(_na) || _l > NETMAP_BUF_SIZE(_na)) {	\
979 		RD(5, "bad addr/len ring %d slot %d idx %d len %d",	\
980 			kring->ring_id, nm_i, slot->buf_idx, len);	\
981 		if (_l > NETMAP_BUF_SIZE(_na))				\
982 			_l = NETMAP_BUF_SIZE(_na);			\
983 	} } while (0)
984 #else /* no debug version */
985 #define	NM_CHECK_ADDR_LEN(_na, _a, _l)	do {				\
986 		if (_l > NETMAP_BUF_SIZE(_na))				\
987 			_l = NETMAP_BUF_SIZE(_na);			\
988 	} while (0)
989 #endif
990 
991 
992 /*---------------------------------------------------------------*/
993 /*
994  * Support routines used by netmap subsystems
995  * (native drivers, VALE, generic, pipes, monitors, ...)
996  */
997 
998 
999 /* common routine for all functions that create a netmap adapter. It performs
1000  * two main tasks:
1001  * - if the na points to an ifp, mark the ifp as netmap capable
1002  *   using na as its native adapter;
1003  * - provide defaults for the setup callbacks and the memory allocator
1004  */
1005 int netmap_attach_common(struct netmap_adapter *);
1006 /* common actions to be performed on netmap adapter destruction */
1007 void netmap_detach_common(struct netmap_adapter *);
1008 /* fill priv->np_[tr]xq{first,last} using the ringid and flags information
1009  * coming from a struct nmreq
1010  */
1011 int netmap_interp_ringid(struct netmap_priv_d *priv, uint16_t ringid, uint32_t flags);
1012 /* update the ring parameters (number and size of tx and rx rings).
1013  * It calls the nm_config callback, if available.
1014  */
1015 int netmap_update_config(struct netmap_adapter *na);
1016 /* create and initialize the common fields of the krings array.
1017  * using the information that must be already available in the na.
1018  * tailroom can be used to request the allocation of additional
1019  * tailroom bytes after the krings array. This is used by
1020  * netmap_vp_adapter's (i.e., VALE ports) to make room for
1021  * leasing-related data structures
1022  */
1023 int netmap_krings_create(struct netmap_adapter *na, u_int tailroom);
1024 /* deletes the kring array of the adapter. The array must have
1025  * been created using netmap_krings_create
1026  */
1027 void netmap_krings_delete(struct netmap_adapter *na);
1028 
1029 /* set the stopped/enabled status of ring
1030  * When stopping, they also wait for all current activity on the ring to
1031  * terminate. The status change is then notified using the na nm_notify
1032  * callback.
1033  */
1034 void netmap_set_txring(struct netmap_adapter *, u_int ring_id, int stopped);
1035 void netmap_set_rxring(struct netmap_adapter *, u_int ring_id, int stopped);
1036 /* set the stopped/enabled status of all rings of the adapter. */
1037 void netmap_set_all_rings(struct netmap_adapter *, int stopped);
1038 /* convenience wrappers for netmap_set_all_rings, used in drivers */
1039 void netmap_disable_all_rings(struct ifnet *);
1040 void netmap_enable_all_rings(struct ifnet *);
1041 
1042 int netmap_rxsync_from_host(struct netmap_adapter *na, struct thread *td, void *pwait);
1043 
1044 struct netmap_if *
1045 netmap_do_regif(struct netmap_priv_d *priv, struct netmap_adapter *na,
1046 	uint16_t ringid, uint32_t flags, int *err);
1047 
1048 
1049 
1050 u_int nm_bound_var(u_int *v, u_int dflt, u_int lo, u_int hi, const char *msg);
1051 int netmap_get_na(struct nmreq *nmr, struct netmap_adapter **na, int create);
1052 int netmap_get_hw_na(struct ifnet *ifp, struct netmap_adapter **na);
1053 
1054 
1055 #ifdef WITH_VALE
1056 /*
1057  * The following bridge-related functions are used by other
1058  * kernel modules.
1059  *
1060  * VALE only supports unicast or broadcast. The lookup
1061  * function can return 0 .. NM_BDG_MAXPORTS-1 for regular ports,
1062  * NM_BDG_MAXPORTS for broadcast, NM_BDG_MAXPORTS+1 for unknown.
1063  * XXX in practice "unknown" might be handled same as broadcast.
1064  */
1065 typedef u_int (*bdg_lookup_fn_t)(struct nm_bdg_fwd *ft, uint8_t *ring_nr,
1066 		const struct netmap_vp_adapter *);
1067 typedef int (*bdg_config_fn_t)(struct nm_ifreq *);
1068 typedef void (*bdg_dtor_fn_t)(const struct netmap_vp_adapter *);
1069 struct netmap_bdg_ops {
1070 	bdg_lookup_fn_t lookup;
1071 	bdg_config_fn_t config;
1072 	bdg_dtor_fn_t	dtor;
1073 };
1074 
1075 u_int netmap_bdg_learning(struct nm_bdg_fwd *ft, uint8_t *dst_ring,
1076 		const struct netmap_vp_adapter *);
1077 
1078 #define	NM_BDG_MAXPORTS		254	/* up to 254 */
1079 #define	NM_BDG_BROADCAST	NM_BDG_MAXPORTS
1080 #define	NM_BDG_NOPORT		(NM_BDG_MAXPORTS+1)
1081 
1082 #define	NM_NAME			"vale"	/* prefix for bridge port name */
1083 
1084 /* these are redefined in case of no VALE support */
1085 int netmap_get_bdg_na(struct nmreq *nmr, struct netmap_adapter **na, int create);
1086 void netmap_init_bridges(void);
1087 int netmap_bdg_ctl(struct nmreq *nmr, struct netmap_bdg_ops *bdg_ops);
1088 int netmap_bdg_config(struct nmreq *nmr);
1089 
1090 #else /* !WITH_VALE */
1091 #define	netmap_get_bdg_na(_1, _2, _3)	0
1092 #define netmap_init_bridges(_1)
1093 #define	netmap_bdg_ctl(_1, _2)	EINVAL
1094 #endif /* !WITH_VALE */
1095 
1096 #ifdef WITH_PIPES
1097 /* max number of pipes per device */
1098 #define NM_MAXPIPES	64	/* XXX how many? */
1099 /* in case of no error, returns the actual number of pipes in nmr->nr_arg1 */
1100 int netmap_pipe_alloc(struct netmap_adapter *, struct nmreq *nmr);
1101 void netmap_pipe_dealloc(struct netmap_adapter *);
1102 int netmap_get_pipe_na(struct nmreq *nmr, struct netmap_adapter **na, int create);
1103 #else /* !WITH_PIPES */
1104 #define NM_MAXPIPES	0
1105 #define netmap_pipe_alloc(_1, _2) 	EOPNOTSUPP
1106 #define netmap_pipe_dealloc(_1)
1107 #define netmap_get_pipe_na(_1, _2, _3)	0
1108 #endif
1109 
1110 #ifdef WITH_MONITOR
1111 int netmap_get_monitor_na(struct nmreq *nmr, struct netmap_adapter **na, int create);
1112 #else
1113 #define netmap_get_monitor_na(_1, _2, _3) 0
1114 #endif
1115 
1116 /* Various prototypes */
1117 int netmap_poll(struct cdev *dev, int events, struct thread *td);
1118 int netmap_init(void);
1119 void netmap_fini(void);
1120 int netmap_get_memory(struct netmap_priv_d* p);
1121 void netmap_dtor(void *data);
1122 int netmap_dtor_locked(struct netmap_priv_d *priv);
1123 
1124 int netmap_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag, struct thread *td);
1125 
1126 /* netmap_adapter creation/destruction */
1127 
1128 // #define NM_DEBUG_PUTGET 1
1129 
1130 #ifdef NM_DEBUG_PUTGET
1131 
1132 #define NM_DBG(f) __##f
1133 
1134 void __netmap_adapter_get(struct netmap_adapter *na);
1135 
1136 #define netmap_adapter_get(na) 				\
1137 	do {						\
1138 		struct netmap_adapter *__na = na;	\
1139 		D("getting %p:%s (%d)", __na, (__na)->name, (__na)->na_refcount);	\
1140 		__netmap_adapter_get(__na);		\
1141 	} while (0)
1142 
1143 int __netmap_adapter_put(struct netmap_adapter *na);
1144 
1145 #define netmap_adapter_put(na)				\
1146 	({						\
1147 		struct netmap_adapter *__na = na;	\
1148 		D("putting %p:%s (%d)", __na, (__na)->name, (__na)->na_refcount);	\
1149 		__netmap_adapter_put(__na);		\
1150 	})
1151 
1152 #else /* !NM_DEBUG_PUTGET */
1153 
1154 #define NM_DBG(f) f
1155 void netmap_adapter_get(struct netmap_adapter *na);
1156 int netmap_adapter_put(struct netmap_adapter *na);
1157 
1158 #endif /* !NM_DEBUG_PUTGET */
1159 
1160 
1161 /*
1162  * module variables
1163  */
1164 #define NETMAP_BUF_BASE(na)	((na)->na_lut[0].vaddr)
1165 #define NETMAP_BUF_SIZE(na)	((na)->na_lut_objsize)
1166 extern int netmap_mitigate;	// XXX not really used
1167 extern int netmap_no_pendintr;
1168 extern int netmap_verbose;	// XXX debugging
1169 enum {                                  /* verbose flags */
1170 	NM_VERB_ON = 1,                 /* generic verbose */
1171 	NM_VERB_HOST = 0x2,             /* verbose host stack */
1172 	NM_VERB_RXSYNC = 0x10,          /* verbose on rxsync/txsync */
1173 	NM_VERB_TXSYNC = 0x20,
1174 	NM_VERB_RXINTR = 0x100,         /* verbose on rx/tx intr (driver) */
1175 	NM_VERB_TXINTR = 0x200,
1176 	NM_VERB_NIC_RXSYNC = 0x1000,    /* verbose on rx/tx intr (driver) */
1177 	NM_VERB_NIC_TXSYNC = 0x2000,
1178 };
1179 
1180 extern int netmap_txsync_retry;
1181 extern int netmap_generic_mit;
1182 extern int netmap_generic_ringsize;
1183 extern int netmap_generic_rings;
1184 
1185 /*
1186  * NA returns a pointer to the struct netmap adapter from the ifp,
1187  * WNA is used to write it.
1188  */
1189 #ifndef WNA
1190 #define	WNA(_ifp)	(_ifp)->if_pspare[0]
1191 #endif
1192 #define	NA(_ifp)	((struct netmap_adapter *)WNA(_ifp))
1193 
1194 /*
1195  * Macros to determine if an interface is netmap capable or netmap enabled.
1196  * See the magic field in struct netmap_adapter.
1197  */
1198 #ifdef __FreeBSD__
1199 /*
1200  * on FreeBSD just use if_capabilities and if_capenable.
1201  */
1202 #define NETMAP_CAPABLE(ifp)	(NA(ifp) &&		\
1203 	(ifp)->if_capabilities & IFCAP_NETMAP )
1204 
1205 #define	NETMAP_SET_CAPABLE(ifp)				\
1206 	(ifp)->if_capabilities |= IFCAP_NETMAP
1207 
1208 #else	/* linux */
1209 
1210 /*
1211  * on linux:
1212  * we check if NA(ifp) is set and its first element has a related
1213  * magic value. The capenable is within the struct netmap_adapter.
1214  */
1215 #define	NETMAP_MAGIC	0x52697a7a
1216 
1217 #define NETMAP_CAPABLE(ifp)	(NA(ifp) &&		\
1218 	((uint32_t)(uintptr_t)NA(ifp) ^ NA(ifp)->magic) == NETMAP_MAGIC )
1219 
1220 #define	NETMAP_SET_CAPABLE(ifp)				\
1221 	NA(ifp)->magic = ((uint32_t)(uintptr_t)NA(ifp)) ^ NETMAP_MAGIC
1222 
1223 #endif	/* linux */
1224 
1225 #ifdef __FreeBSD__
1226 
1227 /* Assigns the device IOMMU domain to an allocator.
1228  * Returns -ENOMEM in case the domain is different */
1229 #define nm_iommu_group_id(dev) (0)
1230 
1231 /* Callback invoked by the dma machinery after a successful dmamap_load */
1232 static void netmap_dmamap_cb(__unused void *arg,
1233     __unused bus_dma_segment_t * segs, __unused int nseg, __unused int error)
1234 {
1235 }
1236 
1237 /* bus_dmamap_load wrapper: call aforementioned function if map != NULL.
1238  * XXX can we do it without a callback ?
1239  */
1240 static inline void
1241 netmap_load_map(struct netmap_adapter *na,
1242 	bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
1243 {
1244 	if (map)
1245 		bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE(na),
1246 		    netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT);
1247 }
1248 
1249 static inline void
1250 netmap_unload_map(struct netmap_adapter *na,
1251         bus_dma_tag_t tag, bus_dmamap_t map)
1252 {
1253 	if (map)
1254 		bus_dmamap_unload(tag, map);
1255 }
1256 
1257 /* update the map when a buffer changes. */
1258 static inline void
1259 netmap_reload_map(struct netmap_adapter *na,
1260 	bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
1261 {
1262 	if (map) {
1263 		bus_dmamap_unload(tag, map);
1264 		bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE(na),
1265 		    netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT);
1266 	}
1267 }
1268 
1269 #else /* linux */
1270 
1271 int nm_iommu_group_id(bus_dma_tag_t dev);
1272 extern size_t     netmap_mem_get_bufsize(struct netmap_mem_d *);
1273 #include <linux/dma-mapping.h>
1274 
1275 static inline void
1276 netmap_load_map(struct netmap_adapter *na,
1277 	bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
1278 {
1279 	if (map) {
1280 		*map = dma_map_single(na->pdev, buf, netmap_mem_get_bufsize(na->nm_mem),
1281 				DMA_BIDIRECTIONAL);
1282 	}
1283 }
1284 
1285 static inline void
1286 netmap_unload_map(struct netmap_adapter *na,
1287 	bus_dma_tag_t tag, bus_dmamap_t map)
1288 {
1289 	u_int sz = netmap_mem_get_bufsize(na->nm_mem);
1290 
1291 	if (*map) {
1292 		dma_unmap_single(na->pdev, *map, sz,
1293 				DMA_BIDIRECTIONAL);
1294 	}
1295 }
1296 
1297 static inline void
1298 netmap_reload_map(struct netmap_adapter *na,
1299 	bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
1300 {
1301 	u_int sz = netmap_mem_get_bufsize(na->nm_mem);
1302 
1303 	if (*map) {
1304 		dma_unmap_single(na->pdev, *map, sz,
1305 				DMA_BIDIRECTIONAL);
1306 	}
1307 
1308 	*map = dma_map_single(na->pdev, buf, sz,
1309 				DMA_BIDIRECTIONAL);
1310 }
1311 
1312 /*
1313  * XXX How do we redefine these functions:
1314  *
1315  * on linux we need
1316  *	dma_map_single(&pdev->dev, virt_addr, len, direction)
1317  *	dma_unmap_single(&adapter->pdev->dev, phys_addr, len, direction
1318  * The len can be implicit (on netmap it is NETMAP_BUF_SIZE)
1319  * unfortunately the direction is not, so we need to change
1320  * something to have a cross API
1321  */
1322 
1323 #if 0
1324 	struct e1000_buffer *buffer_info =  &tx_ring->buffer_info[l];
1325 	/* set time_stamp *before* dma to help avoid a possible race */
1326 	buffer_info->time_stamp = jiffies;
1327 	buffer_info->mapped_as_page = false;
1328 	buffer_info->length = len;
1329 	//buffer_info->next_to_watch = l;
1330 	/* reload dma map */
1331 	dma_unmap_single(&adapter->pdev->dev, buffer_info->dma,
1332 			NETMAP_BUF_SIZE, DMA_TO_DEVICE);
1333 	buffer_info->dma = dma_map_single(&adapter->pdev->dev,
1334 			addr, NETMAP_BUF_SIZE, DMA_TO_DEVICE);
1335 
1336 	if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) {
1337 		D("dma mapping error");
1338 		/* goto dma_error; See e1000_put_txbuf() */
1339 		/* XXX reset */
1340 	}
1341 	tx_desc->buffer_addr = htole64(buffer_info->dma); //XXX
1342 
1343 #endif
1344 
1345 /*
1346  * The bus_dmamap_sync() can be one of wmb() or rmb() depending on direction.
1347  */
1348 #define bus_dmamap_sync(_a, _b, _c)
1349 
1350 #endif /* linux */
1351 
1352 
1353 /*
1354  * functions to map NIC to KRING indexes (n2k) and vice versa (k2n)
1355  */
1356 static inline int
1357 netmap_idx_n2k(struct netmap_kring *kr, int idx)
1358 {
1359 	int n = kr->nkr_num_slots;
1360 	idx += kr->nkr_hwofs;
1361 	if (idx < 0)
1362 		return idx + n;
1363 	else if (idx < n)
1364 		return idx;
1365 	else
1366 		return idx - n;
1367 }
1368 
1369 
1370 static inline int
1371 netmap_idx_k2n(struct netmap_kring *kr, int idx)
1372 {
1373 	int n = kr->nkr_num_slots;
1374 	idx -= kr->nkr_hwofs;
1375 	if (idx < 0)
1376 		return idx + n;
1377 	else if (idx < n)
1378 		return idx;
1379 	else
1380 		return idx - n;
1381 }
1382 
1383 
1384 /* Entries of the look-up table. */
1385 struct lut_entry {
1386 	void *vaddr;		/* virtual address. */
1387 	vm_paddr_t paddr;	/* physical address. */
1388 };
1389 
1390 struct netmap_obj_pool;
1391 
1392 /*
1393  * NMB return the virtual address of a buffer (buffer 0 on bad index)
1394  * PNMB also fills the physical address
1395  */
1396 static inline void *
1397 NMB(struct netmap_adapter *na, struct netmap_slot *slot)
1398 {
1399 	struct lut_entry *lut = na->na_lut;
1400 	uint32_t i = slot->buf_idx;
1401 	return (unlikely(i >= na->na_lut_objtotal)) ?
1402 		lut[0].vaddr : lut[i].vaddr;
1403 }
1404 
1405 static inline void *
1406 PNMB(struct netmap_adapter *na, struct netmap_slot *slot, uint64_t *pp)
1407 {
1408 	uint32_t i = slot->buf_idx;
1409 	struct lut_entry *lut = na->na_lut;
1410 	void *ret = (i >= na->na_lut_objtotal) ? lut[0].vaddr : lut[i].vaddr;
1411 
1412 	*pp = (i >= na->na_lut_objtotal) ? lut[0].paddr : lut[i].paddr;
1413 	return ret;
1414 }
1415 
1416 /* Generic version of NMB, which uses device-specific memory. */
1417 
1418 
1419 
1420 void netmap_txsync_to_host(struct netmap_adapter *na);
1421 
1422 
1423 /*
1424  * Structure associated to each thread which registered an interface.
1425  *
1426  * The first 4 fields of this structure are written by NIOCREGIF and
1427  * read by poll() and NIOC?XSYNC.
1428  *
1429  * There is low contention among writers (a correct user program
1430  * should have none) and among writers and readers, so we use a
1431  * single global lock to protect the structure initialization;
1432  * since initialization involves the allocation of memory,
1433  * we reuse the memory allocator lock.
1434  *
1435  * Read access to the structure is lock free. Readers must check that
1436  * np_nifp is not NULL before using the other fields.
1437  * If np_nifp is NULL initialization has not been performed,
1438  * so they should return an error to userspace.
1439  *
1440  * The ref_done field is used to regulate access to the refcount in the
1441  * memory allocator. The refcount must be incremented at most once for
1442  * each open("/dev/netmap"). The increment is performed by the first
1443  * function that calls netmap_get_memory() (currently called by
1444  * mmap(), NIOCGINFO and NIOCREGIF).
1445  * If the refcount is incremented, it is then decremented when the
1446  * private structure is destroyed.
1447  */
1448 struct netmap_priv_d {
1449 	struct netmap_if * volatile np_nifp;	/* netmap if descriptor. */
1450 
1451 	struct netmap_adapter	*np_na;
1452 	uint32_t	np_flags;	/* from the ioctl */
1453 	u_int		np_txqfirst, np_txqlast; /* range of tx rings to scan */
1454 	u_int		np_rxqfirst, np_rxqlast; /* range of rx rings to scan */
1455 	uint16_t	np_txpoll;	/* XXX and also np_rxpoll ? */
1456 
1457 	struct netmap_mem_d     *np_mref;	/* use with NMG_LOCK held */
1458 	/* np_refcount is only used on FreeBSD */
1459 	int		np_refcount;	/* use with NMG_LOCK held */
1460 
1461 	/* pointers to the selinfo to be used for selrecord.
1462 	 * Either the local or the global one depending on the
1463 	 * number of rings.
1464 	 */
1465 	NM_SELINFO_T *np_rxsi, *np_txsi;
1466 	struct thread	*np_td;		/* kqueue, just debugging */
1467 };
1468 
1469 #ifdef WITH_MONITOR
1470 
1471 struct netmap_monitor_adapter {
1472 	struct netmap_adapter up;
1473 
1474 	struct netmap_priv_d priv;
1475 	uint32_t flags;
1476 };
1477 
1478 #endif /* WITH_MONITOR */
1479 
1480 
1481 /*
1482  * generic netmap emulation for devices that do not have
1483  * native netmap support.
1484  */
1485 int generic_netmap_attach(struct ifnet *ifp);
1486 
1487 int netmap_catch_rx(struct netmap_adapter *na, int intercept);
1488 void generic_rx_handler(struct ifnet *ifp, struct mbuf *m);;
1489 void netmap_catch_tx(struct netmap_generic_adapter *na, int enable);
1490 int generic_xmit_frame(struct ifnet *ifp, struct mbuf *m, void *addr, u_int len, u_int ring_nr);
1491 int generic_find_num_desc(struct ifnet *ifp, u_int *tx, u_int *rx);
1492 void generic_find_num_queues(struct ifnet *ifp, u_int *txq, u_int *rxq);
1493 
1494 //#define RATE_GENERIC  /* Enables communication statistics for generic. */
1495 #ifdef RATE_GENERIC
1496 void generic_rate(int txp, int txs, int txi, int rxp, int rxs, int rxi);
1497 #else
1498 #define generic_rate(txp, txs, txi, rxp, rxs, rxi)
1499 #endif
1500 
1501 /*
1502  * netmap_mitigation API. This is used by the generic adapter
1503  * to reduce the number of interrupt requests/selwakeup
1504  * to clients on incoming packets.
1505  */
1506 void netmap_mitigation_init(struct nm_generic_mit *mit, int idx,
1507                                 struct netmap_adapter *na);
1508 void netmap_mitigation_start(struct nm_generic_mit *mit);
1509 void netmap_mitigation_restart(struct nm_generic_mit *mit);
1510 int netmap_mitigation_active(struct nm_generic_mit *mit);
1511 void netmap_mitigation_cleanup(struct nm_generic_mit *mit);
1512 
1513 
1514 
1515 /* Shared declarations for the VALE switch. */
1516 
1517 /*
1518  * Each transmit queue accumulates a batch of packets into
1519  * a structure before forwarding. Packets to the same
1520  * destination are put in a list using ft_next as a link field.
1521  * ft_frags and ft_next are valid only on the first fragment.
1522  */
1523 struct nm_bdg_fwd {	/* forwarding entry for a bridge */
1524 	void *ft_buf;		/* netmap or indirect buffer */
1525 	uint8_t ft_frags;	/* how many fragments (only on 1st frag) */
1526 	uint8_t _ft_port;	/* dst port (unused) */
1527 	uint16_t ft_flags;	/* flags, e.g. indirect */
1528 	uint16_t ft_len;	/* src fragment len */
1529 	uint16_t ft_next;	/* next packet to same destination */
1530 };
1531 
1532 /* struct 'virtio_net_hdr' from linux. */
1533 struct nm_vnet_hdr {
1534 #define VIRTIO_NET_HDR_F_NEEDS_CSUM     1	/* Use csum_start, csum_offset */
1535 #define VIRTIO_NET_HDR_F_DATA_VALID    2	/* Csum is valid */
1536     uint8_t flags;
1537 #define VIRTIO_NET_HDR_GSO_NONE         0       /* Not a GSO frame */
1538 #define VIRTIO_NET_HDR_GSO_TCPV4        1       /* GSO frame, IPv4 TCP (TSO) */
1539 #define VIRTIO_NET_HDR_GSO_UDP          3       /* GSO frame, IPv4 UDP (UFO) */
1540 #define VIRTIO_NET_HDR_GSO_TCPV6        4       /* GSO frame, IPv6 TCP */
1541 #define VIRTIO_NET_HDR_GSO_ECN          0x80    /* TCP has ECN set */
1542     uint8_t gso_type;
1543     uint16_t hdr_len;
1544     uint16_t gso_size;
1545     uint16_t csum_start;
1546     uint16_t csum_offset;
1547 };
1548 
1549 #define WORST_CASE_GSO_HEADER	(14+40+60)  /* IPv6 + TCP */
1550 
1551 /* Private definitions for IPv4, IPv6, UDP and TCP headers. */
1552 
1553 struct nm_iphdr {
1554 	uint8_t		version_ihl;
1555 	uint8_t		tos;
1556 	uint16_t	tot_len;
1557 	uint16_t	id;
1558 	uint16_t	frag_off;
1559 	uint8_t		ttl;
1560 	uint8_t		protocol;
1561 	uint16_t	check;
1562 	uint32_t	saddr;
1563 	uint32_t	daddr;
1564 	/*The options start here. */
1565 };
1566 
1567 struct nm_tcphdr {
1568 	uint16_t	source;
1569 	uint16_t	dest;
1570 	uint32_t	seq;
1571 	uint32_t	ack_seq;
1572 	uint8_t		doff;  /* Data offset + Reserved */
1573 	uint8_t		flags;
1574 	uint16_t	window;
1575 	uint16_t	check;
1576 	uint16_t	urg_ptr;
1577 };
1578 
1579 struct nm_udphdr {
1580 	uint16_t	source;
1581 	uint16_t	dest;
1582 	uint16_t	len;
1583 	uint16_t	check;
1584 };
1585 
1586 struct nm_ipv6hdr {
1587 	uint8_t		priority_version;
1588 	uint8_t		flow_lbl[3];
1589 
1590 	uint16_t	payload_len;
1591 	uint8_t		nexthdr;
1592 	uint8_t		hop_limit;
1593 
1594 	uint8_t		saddr[16];
1595 	uint8_t		daddr[16];
1596 };
1597 
1598 /* Type used to store a checksum (in host byte order) that hasn't been
1599  * folded yet.
1600  */
1601 #define rawsum_t uint32_t
1602 
1603 rawsum_t nm_csum_raw(uint8_t *data, size_t len, rawsum_t cur_sum);
1604 uint16_t nm_csum_ipv4(struct nm_iphdr *iph);
1605 void nm_csum_tcpudp_ipv4(struct nm_iphdr *iph, void *data,
1606 		      size_t datalen, uint16_t *check);
1607 void nm_csum_tcpudp_ipv6(struct nm_ipv6hdr *ip6h, void *data,
1608 		      size_t datalen, uint16_t *check);
1609 uint16_t nm_csum_fold(rawsum_t cur_sum);
1610 
1611 void bdg_mismatch_datapath(struct netmap_vp_adapter *na,
1612 			   struct netmap_vp_adapter *dst_na,
1613 			   struct nm_bdg_fwd *ft_p, struct netmap_ring *ring,
1614 			   u_int *j, u_int lim, u_int *howmany);
1615 
1616 /* persistent virtual port routines */
1617 int nm_vi_persist(const char *, struct ifnet **);
1618 void nm_vi_detach(struct ifnet *);
1619 void nm_vi_init_index(void);
1620 
1621 #endif /* _NET_NETMAP_KERN_H_ */
1622