1/*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (C) 2013-2016 Vincenzo Maffione
5 * Copyright (C) 2013-2016 Luigi Rizzo
6 * All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 *   1. Redistributions of source code must retain the above copyright
12 *      notice, this list of conditions and the following disclaimer.
13 *   2. Redistributions in binary form must reproduce the above copyright
14 *      notice, this list of conditions and the following disclaimer in the
15 *      documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30/*
31 * This module implements netmap support on top of standard,
32 * unmodified device drivers.
33 *
34 * A NIOCREGIF request is handled here if the device does not
35 * have native support. TX and RX rings are emulated as follows:
36 *
37 * NIOCREGIF
38 *	We preallocate a block of TX mbufs (roughly as many as
39 *	tx descriptors; the number is not critical) to speed up
40 *	operation during transmissions. The refcount on most of
41 *	these buffers is artificially bumped up so we can recycle
42 *	them more easily. Also, the destructor is intercepted
43 *	so we use it as an interrupt notification to wake up
44 *	processes blocked on a poll().
45 *
46 *	For each receive ring we allocate one "struct mbq"
47 *	(an mbuf tailq plus a spinlock). We intercept packets
48 *	(through if_input)
49 *	on the receive path and put them in the mbq from which
50 *	netmap receive routines can grab them.
51 *
52 * TX:
53 *	in the generic_txsync() routine, netmap buffers are copied
54 *	(or linked, in a future) to the preallocated mbufs
55 *	and pushed to the transmit queue. Some of these mbufs
56 *	(those with NS_REPORT, or otherwise every half ring)
57 *	have the refcount=1, others have refcount=2.
58 *	When the destructor is invoked, we take that as
59 *	a notification that all mbufs up to that one in
60 *	the specific ring have been completed, and generate
61 *	the equivalent of a transmit interrupt.
62 *
63 * RX:
64 *
65 */
66
67#ifdef __FreeBSD__
68
69#include <sys/cdefs.h> /* prerequisite */
70#include <sys/types.h>
71#include <sys/errno.h>
72#include <sys/malloc.h>
73#include <sys/lock.h>   /* PROT_EXEC */
74#include <sys/rwlock.h>
75#include <sys/socket.h> /* sockaddrs */
76#include <sys/selinfo.h>
77#include <net/if.h>
78#include <net/if_types.h>
79#include <net/if_var.h>
80#include <machine/bus.h>        /* bus_dmamap_* in netmap_kern.h */
81
82#include <net/netmap.h>
83#include <dev/netmap/netmap_kern.h>
84#include <dev/netmap/netmap_mem2.h>
85
86#define MBUF_RXQ(m)	((m)->m_pkthdr.flowid)
87#define smp_mb()
88
89#elif defined _WIN32
90
91#include "win_glue.h"
92
93#define MBUF_TXQ(m) 	0//((m)->m_pkthdr.flowid)
94#define MBUF_RXQ(m)	    0//((m)->m_pkthdr.flowid)
95#define smp_mb()		//XXX: to be correctly defined
96
97#else /* linux */
98
99#include "bsd_glue.h"
100
101#include <linux/ethtool.h>      /* struct ethtool_ops, get_ringparam */
102#include <linux/hrtimer.h>
103
104static inline struct mbuf *
105nm_os_get_mbuf(struct ifnet *ifp, int len)
106{
107	return alloc_skb(LL_RESERVED_SPACE(ifp) + len +
108			 ifp->needed_tailroom, GFP_ATOMIC);
109}
110
111#endif /* linux */
112
113
114/* Common headers. */
115#include <net/netmap.h>
116#include <dev/netmap/netmap_kern.h>
117#include <dev/netmap/netmap_mem2.h>
118
119
120#define for_each_kring_n(_i, _k, _karr, _n) \
121	for ((_k)=*(_karr), (_i) = 0; (_i) < (_n); (_i)++, (_k) = (_karr)[(_i)])
122
123#define for_each_tx_kring(_i, _k, _na) \
124		for_each_kring_n(_i, _k, (_na)->tx_rings, (_na)->num_tx_rings)
125#define for_each_tx_kring_h(_i, _k, _na) \
126		for_each_kring_n(_i, _k, (_na)->tx_rings, (_na)->num_tx_rings + 1)
127
128#define for_each_rx_kring(_i, _k, _na) \
129		for_each_kring_n(_i, _k, (_na)->rx_rings, (_na)->num_rx_rings)
130#define for_each_rx_kring_h(_i, _k, _na) \
131		for_each_kring_n(_i, _k, (_na)->rx_rings, (_na)->num_rx_rings + 1)
132
133
134/* ======================== PERFORMANCE STATISTICS =========================== */
135
136#ifdef RATE_GENERIC
137#define IFRATE(x) x
138struct rate_stats {
139	unsigned long txpkt;
140	unsigned long txsync;
141	unsigned long txirq;
142	unsigned long txrepl;
143	unsigned long txdrop;
144	unsigned long rxpkt;
145	unsigned long rxirq;
146	unsigned long rxsync;
147};
148
149struct rate_context {
150	unsigned refcount;
151	struct timer_list timer;
152	struct rate_stats new;
153	struct rate_stats old;
154};
155
156#define RATE_PRINTK(_NAME_) \
157	printk( #_NAME_ " = %lu Hz\n", (cur._NAME_ - ctx->old._NAME_)/RATE_PERIOD);
158#define RATE_PERIOD  2
159static void rate_callback(unsigned long arg)
160{
161	struct rate_context * ctx = (struct rate_context *)arg;
162	struct rate_stats cur = ctx->new;
163	int r;
164
165	RATE_PRINTK(txpkt);
166	RATE_PRINTK(txsync);
167	RATE_PRINTK(txirq);
168	RATE_PRINTK(txrepl);
169	RATE_PRINTK(txdrop);
170	RATE_PRINTK(rxpkt);
171	RATE_PRINTK(rxsync);
172	RATE_PRINTK(rxirq);
173	printk("\n");
174
175	ctx->old = cur;
176	r = mod_timer(&ctx->timer, jiffies +
177			msecs_to_jiffies(RATE_PERIOD * 1000));
178	if (unlikely(r))
179		nm_prerr("mod_timer() failed");
180}
181
182static struct rate_context rate_ctx;
183
184void generic_rate(int txp, int txs, int txi, int rxp, int rxs, int rxi)
185{
186	if (txp) rate_ctx.new.txpkt++;
187	if (txs) rate_ctx.new.txsync++;
188	if (txi) rate_ctx.new.txirq++;
189	if (rxp) rate_ctx.new.rxpkt++;
190	if (rxs) rate_ctx.new.rxsync++;
191	if (rxi) rate_ctx.new.rxirq++;
192}
193
194#else /* !RATE */
195#define IFRATE(x)
196#endif /* !RATE */
197
198
199/* ========== GENERIC (EMULATED) NETMAP ADAPTER SUPPORT ============= */
200
201/*
202 * Wrapper used by the generic adapter layer to notify
203 * the poller threads. Differently from netmap_rx_irq(), we check
204 * only NAF_NETMAP_ON instead of NAF_NATIVE_ON to enable the irq.
205 */
206void
207netmap_generic_irq(struct netmap_adapter *na, u_int q, u_int *work_done)
208{
209	if (unlikely(!nm_netmap_on(na)))
210		return;
211
212	netmap_common_irq(na, q, work_done);
213#ifdef RATE_GENERIC
214	if (work_done)
215		rate_ctx.new.rxirq++;
216	else
217		rate_ctx.new.txirq++;
218#endif  /* RATE_GENERIC */
219}
220
221static int
222generic_netmap_unregister(struct netmap_adapter *na)
223{
224	struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
225	struct netmap_kring *kring = NULL;
226	int i, r;
227
228	if (na->active_fds == 0) {
229		na->na_flags &= ~NAF_NETMAP_ON;
230
231		/* Stop intercepting packets on the RX path. */
232		nm_os_catch_rx(gna, 0);
233
234		/* Release packet steering control. */
235		nm_os_catch_tx(gna, 0);
236	}
237
238	netmap_krings_mode_commit(na, /*onoff=*/0);
239
240	for_each_rx_kring(r, kring, na) {
241		/* Free the mbufs still pending in the RX queues,
242		 * that did not end up into the corresponding netmap
243		 * RX rings. */
244		mbq_safe_purge(&kring->rx_queue);
245		nm_os_mitigation_cleanup(&gna->mit[r]);
246	}
247
248	/* Decrement reference counter for the mbufs in the
249	 * TX pools. These mbufs can be still pending in drivers,
250	 * (e.g. this happens with virtio-net driver, which
251	 * does lazy reclaiming of transmitted mbufs). */
252	for_each_tx_kring(r, kring, na) {
253		/* We must remove the destructor on the TX event,
254		 * because the destructor invokes netmap code, and
255		 * the netmap module may disappear before the
256		 * TX event is consumed. */
257		mtx_lock_spin(&kring->tx_event_lock);
258		if (kring->tx_event) {
259			SET_MBUF_DESTRUCTOR(kring->tx_event, NULL, NULL);
260		}
261		kring->tx_event = NULL;
262		mtx_unlock_spin(&kring->tx_event_lock);
263	}
264
265	if (na->active_fds == 0) {
266		nm_os_free(gna->mit);
267
268		for_each_rx_kring(r, kring, na) {
269			mbq_safe_fini(&kring->rx_queue);
270		}
271
272		for_each_tx_kring(r, kring, na) {
273			callout_drain(&kring->tx_event_callout);
274
275			if (kring->tx_pool == NULL) {
276				continue;
277			}
278
279			for (i=0; i<na->num_tx_desc; i++) {
280				if (kring->tx_pool[i]) {
281					m_free(kring->tx_pool[i]);
282					kring->tx_pool[i] = NULL;
283				}
284			}
285			mtx_destroy(&kring->tx_event_lock);
286			nm_os_free(kring->tx_pool);
287			kring->tx_pool = NULL;
288		}
289
290#ifdef RATE_GENERIC
291		if (--rate_ctx.refcount == 0) {
292			nm_prinf("del_timer()");
293			del_timer(&rate_ctx.timer);
294		}
295#endif
296		nm_prinf("Emulated adapter for %s deactivated", na->name);
297	}
298
299	return 0;
300}
301
302/* Enable/disable netmap mode for a generic network interface. */
303static int
304generic_netmap_register(struct netmap_adapter *na, int enable)
305{
306	struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
307	struct netmap_kring *kring = NULL;
308	int error;
309	int i, r;
310
311	if (!na) {
312		return EINVAL;
313	}
314
315	if (!enable) {
316		/* This is actually an unregif. */
317		return generic_netmap_unregister(na);
318	}
319
320	if (na->active_fds == 0) {
321		nm_prinf("Emulated adapter for %s activated", na->name);
322		/* Do all memory allocations when (na->active_fds == 0), to
323		 * simplify error management. */
324
325		/* Allocate memory for mitigation support on all the rx queues. */
326		gna->mit = nm_os_malloc(na->num_rx_rings * sizeof(struct nm_generic_mit));
327		if (!gna->mit) {
328			nm_prerr("mitigation allocation failed");
329			error = ENOMEM;
330			goto out;
331		}
332
333		for_each_rx_kring(r, kring, na) {
334			/* Init mitigation support. */
335			nm_os_mitigation_init(&gna->mit[r], r, na);
336
337			/* Initialize the rx queue, as generic_rx_handler() can
338			 * be called as soon as nm_os_catch_rx() returns.
339			 */
340			mbq_safe_init(&kring->rx_queue);
341		}
342
343		/*
344		 * Prepare mbuf pools (parallel to the tx rings), for packet
345		 * transmission. Don't preallocate the mbufs here, it's simpler
346		 * to leave this task to txsync.
347		 */
348		for_each_tx_kring(r, kring, na) {
349			kring->tx_pool = NULL;
350		}
351		for_each_tx_kring(r, kring, na) {
352			kring->tx_pool =
353				nm_os_malloc(na->num_tx_desc * sizeof(struct mbuf *));
354			if (!kring->tx_pool) {
355				nm_prerr("tx_pool allocation failed");
356				error = ENOMEM;
357				goto free_tx_pools;
358			}
359			mtx_init(&kring->tx_event_lock, "tx_event_lock",
360				 NULL, MTX_SPIN);
361			callout_init_mtx(&kring->tx_event_callout,
362					 &kring->tx_event_lock,
363					 CALLOUT_RETURNUNLOCKED);
364		}
365	}
366
367	netmap_krings_mode_commit(na, /*onoff=*/1);
368
369	for_each_tx_kring(r, kring, na) {
370		/* Initialize tx_pool and tx_event. */
371		for (i=0; i<na->num_tx_desc; i++) {
372			kring->tx_pool[i] = NULL;
373		}
374
375		kring->tx_event = NULL;
376	}
377
378	if (na->active_fds == 0) {
379		/* Prepare to intercept incoming traffic. */
380		error = nm_os_catch_rx(gna, 1);
381		if (error) {
382			nm_prerr("nm_os_catch_rx(1) failed (%d)", error);
383			goto free_tx_pools;
384		}
385
386		/* Let netmap control the packet steering. */
387		error = nm_os_catch_tx(gna, 1);
388		if (error) {
389			nm_prerr("nm_os_catch_tx(1) failed (%d)", error);
390			goto catch_rx;
391		}
392
393		na->na_flags |= NAF_NETMAP_ON;
394
395#ifdef RATE_GENERIC
396		if (rate_ctx.refcount == 0) {
397			nm_prinf("setup_timer()");
398			memset(&rate_ctx, 0, sizeof(rate_ctx));
399			setup_timer(&rate_ctx.timer, &rate_callback, (unsigned long)&rate_ctx);
400			if (mod_timer(&rate_ctx.timer, jiffies + msecs_to_jiffies(1500))) {
401				nm_prerr("Error: mod_timer()");
402			}
403		}
404		rate_ctx.refcount++;
405#endif /* RATE */
406	}
407
408	return 0;
409
410	/* Here (na->active_fds == 0) holds. */
411catch_rx:
412	nm_os_catch_rx(gna, 0);
413free_tx_pools:
414	for_each_tx_kring(r, kring, na) {
415		mtx_destroy(&kring->tx_event_lock);
416		if (kring->tx_pool == NULL) {
417			continue;
418		}
419		nm_os_free(kring->tx_pool);
420		kring->tx_pool = NULL;
421	}
422	for_each_rx_kring(r, kring, na) {
423		mbq_safe_fini(&kring->rx_queue);
424	}
425	nm_os_free(gna->mit);
426out:
427
428	return error;
429}
430
431/*
432 * Callback invoked when the device driver frees an mbuf used
433 * by netmap to transmit a packet. This usually happens when
434 * the NIC notifies the driver that transmission is completed.
435 */
436static void
437generic_mbuf_dtor(struct mbuf *m)
438{
439	struct netmap_adapter *na = GEN_TX_MBUF_NA(m);
440	struct netmap_kring *kring;
441	unsigned int r = MBUF_TXQ(m);
442	unsigned int r_orig = r;
443
444	if (unlikely(!nm_netmap_on(na) || r >= na->num_tx_rings)) {
445		nm_prerr("Error: no netmap adapter on device %p",
446		  GEN_TX_MBUF_IFP(m));
447		return;
448	}
449
450	/*
451	 * First, clear the event mbuf.
452	 * In principle, the event 'm' should match the one stored
453	 * on ring 'r'. However we check it explicitly to stay
454	 * safe against lower layers (qdisc, driver, etc.) changing
455	 * MBUF_TXQ(m) under our feet. If the match is not found
456	 * on 'r', we try to see if it belongs to some other ring.
457	 */
458	for (;;) {
459		bool match = false;
460
461		kring = na->tx_rings[r];
462		mtx_lock_spin(&kring->tx_event_lock);
463
464		/*
465		 * The netmap destructor can be called between us getting the
466		 * reference and taking the lock, in that case the ring
467		 * reference won't be valid. The destructor will free this mbuf
468		 * so we can stop here.
469		 */
470		if (GEN_TX_MBUF_NA(m) == NULL) {
471			mtx_unlock_spin(&kring->tx_event_lock);
472			return;
473		}
474
475		if (kring->tx_event == m) {
476			kring->tx_event = NULL;
477			match = true;
478		}
479		mtx_unlock_spin(&kring->tx_event_lock);
480
481		if (match) {
482			if (r != r_orig) {
483				nm_prlim(1, "event %p migrated: ring %u --> %u",
484				      m, r_orig, r);
485			}
486			break;
487		}
488
489		if (++r == na->num_tx_rings) r = 0;
490
491		if (r == r_orig) {
492#ifndef __FreeBSD__
493			/*
494			 * On FreeBSD this situation can arise if the tx_event
495			 * callout handler cleared a stuck packet.
496			 */
497			nm_prlim(1, "Cannot match event %p", m);
498#endif
499			nm_generic_mbuf_dtor(m);
500			return;
501		}
502	}
503
504	/* Second, wake up clients. They will reclaim the event through
505	 * txsync. */
506	netmap_generic_irq(na, r, NULL);
507	nm_generic_mbuf_dtor(m);
508}
509
510/* Record completed transmissions and update hwtail.
511 *
512 * The oldest tx buffer not yet completed is at nr_hwtail + 1,
513 * nr_hwcur is the first unsent buffer.
514 */
515static u_int
516generic_netmap_tx_clean(struct netmap_kring *kring, int txqdisc)
517{
518	u_int const lim = kring->nkr_num_slots - 1;
519	u_int nm_i = nm_next(kring->nr_hwtail, lim);
520	u_int hwcur = kring->nr_hwcur;
521	u_int n = 0;
522	struct mbuf **tx_pool = kring->tx_pool;
523
524	nm_prdis("hwcur = %d, hwtail = %d", kring->nr_hwcur, kring->nr_hwtail);
525
526	while (nm_i != hwcur) { /* buffers not completed */
527		struct mbuf *m = tx_pool[nm_i];
528
529		if (txqdisc) {
530			if (m == NULL) {
531				/* Nothing to do, this is going
532				 * to be replenished. */
533				nm_prlim(3, "Is this happening?");
534
535			} else if (MBUF_QUEUED(m)) {
536				break; /* Not dequeued yet. */
537
538			} else if (MBUF_REFCNT(m) != 1) {
539				/* This mbuf has been dequeued but is still busy
540				 * (refcount is 2).
541				 * Leave it to the driver and replenish. */
542				m_free(m);
543				tx_pool[nm_i] = NULL;
544			}
545
546		} else {
547			if (unlikely(m == NULL)) {
548				int event_consumed;
549
550				/* This slot was used to place an event. */
551				mtx_lock_spin(&kring->tx_event_lock);
552				event_consumed = (kring->tx_event == NULL);
553				mtx_unlock_spin(&kring->tx_event_lock);
554				if (!event_consumed) {
555					/* The event has not been consumed yet,
556					 * still busy in the driver. */
557					break;
558				}
559				/* The event has been consumed, we can go
560				 * ahead. */
561			} else if (MBUF_REFCNT(m) != 1) {
562				/* This mbuf is still busy: its refcnt is 2. */
563				break;
564			}
565		}
566
567		n++;
568		nm_i = nm_next(nm_i, lim);
569	}
570	kring->nr_hwtail = nm_prev(nm_i, lim);
571	nm_prdis("tx completed [%d] -> hwtail %d", n, kring->nr_hwtail);
572
573	return n;
574}
575
576/* Compute a slot index in the middle between inf and sup. */
577static inline u_int
578ring_middle(u_int inf, u_int sup, u_int lim)
579{
580	u_int n = lim + 1;
581	u_int e;
582
583	if (sup >= inf) {
584		e = (sup + inf) / 2;
585	} else { /* wrap around */
586		e = (sup + n + inf) / 2;
587		if (e >= n) {
588			e -= n;
589		}
590	}
591
592	if (unlikely(e >= n)) {
593		nm_prerr("This cannot happen");
594		e = 0;
595	}
596
597	return e;
598}
599
600#ifdef __FreeBSD__
601static void
602generic_tx_callout(void *arg)
603{
604	struct netmap_kring *kring = arg;
605
606	kring->tx_event = NULL;
607	mtx_unlock_spin(&kring->tx_event_lock);
608	netmap_generic_irq(kring->na, kring->ring_id, NULL);
609}
610#endif
611
612static void
613generic_set_tx_event(struct netmap_kring *kring, u_int hwcur)
614{
615	u_int lim = kring->nkr_num_slots - 1;
616	struct mbuf *m;
617	u_int e;
618	u_int ntc = nm_next(kring->nr_hwtail, lim); /* next to clean */
619
620	if (ntc == hwcur) {
621		return; /* all buffers are free */
622	}
623
624	/*
625	 * We have pending packets in the driver between hwtail+1
626	 * and hwcur, and we have to chose one of these slot to
627	 * generate a notification.
628	 * There is a race but this is only called within txsync which
629	 * does a double check.
630	 */
631#if 0
632	/* Choose a slot in the middle, so that we don't risk ending
633	 * up in a situation where the client continuously wake up,
634	 * fills one or a few TX slots and go to sleep again. */
635	e = ring_middle(ntc, hwcur, lim);
636#else
637	/* Choose the first pending slot, to be safe against driver
638	 * reordering mbuf transmissions. */
639	e = ntc;
640#endif
641
642	m = kring->tx_pool[e];
643	if (m == NULL) {
644		/* An event is already in place. */
645		return;
646	}
647
648	mtx_lock_spin(&kring->tx_event_lock);
649	if (kring->tx_event) {
650		/* An event is already in place. */
651		mtx_unlock_spin(&kring->tx_event_lock);
652		return;
653	}
654
655	SET_MBUF_DESTRUCTOR(m, generic_mbuf_dtor, kring->na);
656
657	kring->tx_event = m;
658#ifdef __FreeBSD__
659	/*
660	 * Handle the possibility that the transmitted buffer isn't reclaimed
661	 * within a bounded period of time.  This can arise when transmitting
662	 * out of multiple ports via a lagg or bridge interface, since the
663	 * member ports may legitimately only free transmitted buffers in
664	 * batches.
665	 *
666	 * The callout handler clears the stuck packet from the ring, allowing
667	 * transmission to proceed.  In the common case we let
668	 * generic_mbuf_dtor() unstick the ring, allowing mbufs to be
669	 * reused most of the time.
670	 */
671	callout_reset_sbt_curcpu(&kring->tx_event_callout, SBT_1MS, 0,
672	    generic_tx_callout, kring, 0);
673#endif
674	mtx_unlock_spin(&kring->tx_event_lock);
675
676	kring->tx_pool[e] = NULL;
677
678	nm_prdis("Request Event at %d mbuf %p refcnt %d", e, m, m ? MBUF_REFCNT(m) : -2 );
679
680	/* Decrement the refcount. This will free it if we lose the race
681	 * with the driver. */
682	m_free(m);
683}
684
685/*
686 * generic_netmap_txsync() transforms netmap buffers into mbufs
687 * and passes them to the standard device driver
688 * (ndo_start_xmit() or ifp->if_transmit() ).
689 * On linux this is not done directly, but using dev_queue_xmit(),
690 * since it implements the TX flow control (and takes some locks).
691 */
692static int
693generic_netmap_txsync(struct netmap_kring *kring, int flags)
694{
695	struct netmap_adapter *na = kring->na;
696	struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
697	if_t ifp = na->ifp;
698	struct netmap_ring *ring = kring->ring;
699	u_int nm_i;	/* index into the netmap ring */ // j
700	u_int const lim = kring->nkr_num_slots - 1;
701	u_int const head = kring->rhead;
702	u_int ring_nr = kring->ring_id;
703
704	IFRATE(rate_ctx.new.txsync++);
705
706	rmb();
707
708	/*
709	 * First part: process new packets to send.
710	 */
711	nm_i = kring->nr_hwcur;
712	if (nm_i != head) {	/* we have new packets to send */
713		struct nm_os_gen_arg a;
714		u_int event = -1;
715#ifdef __FreeBSD__
716		struct epoch_tracker et;
717
718		NET_EPOCH_ENTER(et);
719#endif
720
721		if (gna->txqdisc && nm_kr_txempty(kring)) {
722			/* In txqdisc mode, we ask for a delayed notification,
723			 * but only when cur == hwtail, which means that the
724			 * client is going to block. */
725			event = ring_middle(nm_i, head, lim);
726			nm_prdis("Place txqdisc event (hwcur=%u,event=%u,"
727			      "head=%u,hwtail=%u)", nm_i, event, head,
728			      kring->nr_hwtail);
729		}
730
731		a.ifp = ifp;
732		a.ring_nr = ring_nr;
733		a.head = a.tail = NULL;
734
735		while (nm_i != head) {
736			struct netmap_slot *slot = &ring->slot[nm_i];
737			u_int len = slot->len;
738			void *addr = NMB(na, slot);
739			/* device-specific */
740			struct mbuf *m;
741			int tx_ret;
742
743			NM_CHECK_ADDR_LEN(na, addr, len);
744
745			/* Tale a mbuf from the tx pool (replenishing the pool
746			 * entry if necessary) and copy in the user packet. */
747			m = kring->tx_pool[nm_i];
748			if (unlikely(m == NULL)) {
749				kring->tx_pool[nm_i] = m =
750					nm_os_get_mbuf(ifp, NETMAP_BUF_SIZE(na));
751				if (m == NULL) {
752					nm_prlim(2, "Failed to replenish mbuf");
753					/* Here we could schedule a timer which
754					 * retries to replenish after a while,
755					 * and notifies the client when it
756					 * manages to replenish some slots. In
757					 * any case we break early to avoid
758					 * crashes. */
759					break;
760				}
761				IFRATE(rate_ctx.new.txrepl++);
762			} else {
763				nm_os_mbuf_reinit(m);
764			}
765
766			a.m = m;
767			a.addr = addr;
768			a.len = len;
769			a.qevent = (nm_i == event);
770			/* When not in txqdisc mode, we should ask
771			 * notifications when NS_REPORT is set, or roughly
772			 * every half ring. To optimize this, we set a
773			 * notification event when the client runs out of
774			 * TX ring space, or when transmission fails. In
775			 * the latter case we also break early.
776			 */
777			tx_ret = nm_os_generic_xmit_frame(&a);
778			if (unlikely(tx_ret)) {
779				if (!gna->txqdisc) {
780					/*
781					 * No room for this mbuf in the device driver.
782					 * Request a notification FOR A PREVIOUS MBUF,
783					 * then call generic_netmap_tx_clean(kring) to do the
784					 * double check and see if we can free more buffers.
785					 * If there is space continue, else break;
786					 * NOTE: the double check is necessary if the problem
787					 * occurs in the txsync call after selrecord().
788					 * Also, we need some way to tell the caller that not
789					 * all buffers were queued onto the device (this was
790					 * not a problem with native netmap driver where space
791					 * is preallocated). The bridge has a similar problem
792					 * and we solve it there by dropping the excess packets.
793					 */
794					generic_set_tx_event(kring, nm_i);
795					if (generic_netmap_tx_clean(kring, gna->txqdisc)) {
796						/* space now available */
797						continue;
798					} else {
799						break;
800					}
801				}
802
803				/* In txqdisc mode, the netmap-aware qdisc
804				 * queue has the same length as the number of
805				 * netmap slots (N). Since tail is advanced
806				 * only when packets are dequeued, qdisc
807				 * queue overrun cannot happen, so
808				 * nm_os_generic_xmit_frame() did not fail
809				 * because of that.
810				 * However, packets can be dropped because
811				 * carrier is off, or because our qdisc is
812				 * being deactivated, or possibly for other
813				 * reasons. In these cases, we just let the
814				 * packet to be dropped. */
815				IFRATE(rate_ctx.new.txdrop++);
816			}
817
818			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
819			nm_i = nm_next(nm_i, lim);
820			IFRATE(rate_ctx.new.txpkt++);
821		}
822		if (a.head != NULL) {
823			a.addr = NULL;
824			nm_os_generic_xmit_frame(&a);
825		}
826		/* Update hwcur to the next slot to transmit. Here nm_i
827		 * is not necessarily head, we could break early. */
828		kring->nr_hwcur = nm_i;
829
830#ifdef __FreeBSD__
831		NET_EPOCH_EXIT(et);
832#endif
833	}
834
835	if (!gna->txqdisc && (flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring))) {
836		/* No more available slots? Set a notification event
837		 * on a netmap slot that will be cleaned in the future.
838		 * No doublecheck is performed, since txsync() will be
839		 * called twice by netmap_poll().
840		 */
841		generic_set_tx_event(kring, nm_i);
842	}
843
844	/*
845	 * Second, reclaim completed buffers
846	 */
847	generic_netmap_tx_clean(kring, gna->txqdisc);
848
849	return 0;
850}
851
852
853/*
854 * This handler is registered (through nm_os_catch_rx())
855 * within the attached network interface
856 * in the RX subsystem, so that every mbuf passed up by
857 * the driver can be stolen to the network stack.
858 * Stolen packets are put in a queue where the
859 * generic_netmap_rxsync() callback can extract them.
860 * Returns 1 if the packet was stolen, 0 otherwise.
861 */
862int
863generic_rx_handler(if_t ifp, struct mbuf *m)
864{
865	struct netmap_adapter *na = NA(ifp);
866	struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
867	struct netmap_kring *kring;
868	u_int work_done;
869	u_int r = MBUF_RXQ(m); /* receive ring number */
870
871	if (r >= na->num_rx_rings) {
872		r = r % na->num_rx_rings;
873	}
874
875	kring = na->rx_rings[r];
876
877	if (kring->nr_mode == NKR_NETMAP_OFF) {
878		/* We must not intercept this mbuf. */
879		return 0;
880	}
881
882	/* limit the size of the queue */
883	if (unlikely(!gna->rxsg && MBUF_LEN(m) > NETMAP_BUF_SIZE(na))) {
884		/* This may happen when GRO/LRO features are enabled for
885		 * the NIC driver when the generic adapter does not
886		 * support RX scatter-gather. */
887		nm_prlim(2, "Warning: driver pushed up big packet "
888				"(size=%d)", (int)MBUF_LEN(m));
889		if_inc_counter(ifp, IFCOUNTER_IQDROPS, 1);
890		m_freem(m);
891	} else if (unlikely(mbq_len(&kring->rx_queue) > na->num_rx_desc)) {
892		if_inc_counter(ifp, IFCOUNTER_IQDROPS, 1);
893		m_freem(m);
894	} else {
895		mbq_safe_enqueue(&kring->rx_queue, m);
896	}
897
898	if (netmap_generic_mit < 32768) {
899		/* no rx mitigation, pass notification up */
900		netmap_generic_irq(na, r, &work_done);
901	} else {
902		/* same as send combining, filter notification if there is a
903		 * pending timer, otherwise pass it up and start a timer.
904		 */
905		if (likely(nm_os_mitigation_active(&gna->mit[r]))) {
906			/* Record that there is some pending work. */
907			gna->mit[r].mit_pending = 1;
908		} else {
909			netmap_generic_irq(na, r, &work_done);
910			nm_os_mitigation_start(&gna->mit[r]);
911		}
912	}
913
914	/* We have intercepted the mbuf. */
915	return 1;
916}
917
918/*
919 * generic_netmap_rxsync() extracts mbufs from the queue filled by
920 * generic_netmap_rx_handler() and puts their content in the netmap
921 * receive ring.
922 * Access must be protected because the rx handler is asynchronous,
923 */
924static int
925generic_netmap_rxsync(struct netmap_kring *kring, int flags)
926{
927	struct netmap_ring *ring = kring->ring;
928	struct netmap_adapter *na = kring->na;
929	u_int nm_i;	/* index into the netmap ring */ //j,
930	u_int n;
931	u_int const lim = kring->nkr_num_slots - 1;
932	u_int const head = kring->rhead;
933	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
934
935	/* Adapter-specific variables. */
936	u_int nm_buf_len = NETMAP_BUF_SIZE(na);
937	struct mbq tmpq;
938	struct mbuf *m;
939	int avail; /* in bytes */
940	int mlen;
941	int copy;
942
943	if (head > lim)
944		return netmap_ring_reinit(kring);
945
946	IFRATE(rate_ctx.new.rxsync++);
947
948	/*
949	 * First part: skip past packets that userspace has released.
950	 * This can possibly make room for the second part.
951	 */
952	nm_i = kring->nr_hwcur;
953	if (nm_i != head) {
954		/* Userspace has released some packets. */
955		for (n = 0; nm_i != head; n++) {
956			struct netmap_slot *slot = &ring->slot[nm_i];
957
958			slot->flags &= ~NS_BUF_CHANGED;
959			nm_i = nm_next(nm_i, lim);
960		}
961		kring->nr_hwcur = head;
962	}
963
964	/*
965	 * Second part: import newly received packets.
966	 */
967	if (!netmap_no_pendintr && !force_update) {
968		return 0;
969	}
970
971	nm_i = kring->nr_hwtail; /* First empty slot in the receive ring. */
972
973	/* Compute the available space (in bytes) in this netmap ring.
974	 * The first slot that is not considered in is the one before
975	 * nr_hwcur. */
976
977	avail = nm_prev(kring->nr_hwcur, lim) - nm_i;
978	if (avail < 0)
979		avail += lim + 1;
980	avail *= nm_buf_len;
981
982	/* First pass: While holding the lock on the RX mbuf queue,
983	 * extract as many mbufs as they fit the available space,
984	 * and put them in a temporary queue.
985	 * To avoid performing a per-mbuf division (mlen / nm_buf_len) to
986	 * to update avail, we do the update in a while loop that we
987	 * also use to set the RX slots, but without performing the copy. */
988	mbq_init(&tmpq);
989	mbq_lock(&kring->rx_queue);
990	for (n = 0;; n++) {
991		m = mbq_peek(&kring->rx_queue);
992		if (!m) {
993			/* No more packets from the driver. */
994			break;
995		}
996
997		mlen = MBUF_LEN(m);
998		if (mlen > avail) {
999			/* No more space in the ring. */
1000			break;
1001		}
1002
1003		mbq_dequeue(&kring->rx_queue);
1004
1005		while (mlen) {
1006			copy = nm_buf_len;
1007			if (mlen < copy) {
1008				copy = mlen;
1009			}
1010			mlen -= copy;
1011			avail -= nm_buf_len;
1012
1013			ring->slot[nm_i].len = copy;
1014			ring->slot[nm_i].flags = (mlen ? NS_MOREFRAG : 0);
1015			nm_i = nm_next(nm_i, lim);
1016		}
1017
1018		mbq_enqueue(&tmpq, m);
1019	}
1020	mbq_unlock(&kring->rx_queue);
1021
1022	/* Second pass: Drain the temporary queue, going over the used RX slots,
1023	 * and perform the copy out of the RX queue lock. */
1024	nm_i = kring->nr_hwtail;
1025
1026	for (;;) {
1027		void *nmaddr;
1028		int ofs = 0;
1029		int morefrag;
1030
1031		m = mbq_dequeue(&tmpq);
1032		if (!m)	{
1033			break;
1034		}
1035
1036		do {
1037			nmaddr = NMB(na, &ring->slot[nm_i]);
1038			/* We only check the address here on generic rx rings. */
1039			if (nmaddr == NETMAP_BUF_BASE(na)) { /* Bad buffer */
1040				m_freem(m);
1041				mbq_purge(&tmpq);
1042				mbq_fini(&tmpq);
1043				return netmap_ring_reinit(kring);
1044			}
1045
1046			copy = ring->slot[nm_i].len;
1047			m_copydata(m, ofs, copy, nmaddr);
1048			ofs += copy;
1049			morefrag = ring->slot[nm_i].flags & NS_MOREFRAG;
1050			nm_i = nm_next(nm_i, lim);
1051		} while (morefrag);
1052
1053		m_freem(m);
1054	}
1055
1056	mbq_fini(&tmpq);
1057
1058	if (n) {
1059		kring->nr_hwtail = nm_i;
1060		IFRATE(rate_ctx.new.rxpkt += n);
1061	}
1062	kring->nr_kflags &= ~NKR_PENDINTR;
1063
1064	return 0;
1065}
1066
1067static void
1068generic_netmap_dtor(struct netmap_adapter *na)
1069{
1070	struct netmap_generic_adapter *gna = (struct netmap_generic_adapter*)na;
1071	if_t ifp = netmap_generic_getifp(gna);
1072	struct netmap_adapter *prev_na = gna->prev;
1073
1074	if (prev_na != NULL) {
1075		netmap_adapter_put(prev_na);
1076		if (nm_iszombie(na)) {
1077		        /*
1078		         * The driver has been removed without releasing
1079		         * the reference so we need to do it here.
1080		         */
1081		        netmap_adapter_put(prev_na);
1082		}
1083		nm_prinf("Native netmap adapter for %s restored", prev_na->name);
1084	}
1085	NM_RESTORE_NA(ifp, prev_na);
1086	na->ifp = NULL;
1087	nm_prinf("Emulated netmap adapter for %s destroyed", na->name);
1088}
1089
1090int
1091na_is_generic(struct netmap_adapter *na)
1092{
1093	return na->nm_register == generic_netmap_register;
1094}
1095
1096/*
1097 * generic_netmap_attach() makes it possible to use netmap on
1098 * a device without native netmap support.
1099 * This is less performant than native support but potentially
1100 * faster than raw sockets or similar schemes.
1101 *
1102 * In this "emulated" mode, netmap rings do not necessarily
1103 * have the same size as those in the NIC. We use a default
1104 * value and possibly override it if the OS has ways to fetch the
1105 * actual configuration.
1106 */
1107int
1108generic_netmap_attach(if_t ifp)
1109{
1110	struct netmap_adapter *na;
1111	struct netmap_generic_adapter *gna;
1112	int retval;
1113	u_int num_tx_desc, num_rx_desc;
1114
1115#ifdef __FreeBSD__
1116	if (if_gettype(ifp) == IFT_LOOP) {
1117		nm_prerr("if_loop is not supported by %s", __func__);
1118		return EINVAL;
1119	}
1120#endif
1121
1122	if (NM_NA_CLASH(ifp)) {
1123		/* If NA(ifp) is not null but there is no valid netmap
1124		 * adapter it means that someone else is using the same
1125		 * pointer (e.g. ax25_ptr on linux). This happens for
1126		 * instance when also PF_RING is in use. */
1127		nm_prerr("Error: netmap adapter hook is busy");
1128		return EBUSY;
1129	}
1130
1131	num_tx_desc = num_rx_desc = netmap_generic_ringsize; /* starting point */
1132
1133	nm_os_generic_find_num_desc(ifp, &num_tx_desc, &num_rx_desc); /* ignore errors */
1134	if (num_tx_desc == 0 || num_rx_desc == 0) {
1135		nm_prerr("Device has no hw slots (tx %u, rx %u)", num_tx_desc, num_rx_desc);
1136		return EINVAL;
1137	}
1138
1139	gna = nm_os_malloc(sizeof(*gna));
1140	if (gna == NULL) {
1141		nm_prerr("no memory on attach, give up");
1142		return ENOMEM;
1143	}
1144	na = (struct netmap_adapter *)gna;
1145	strlcpy(na->name, if_name(ifp), sizeof(na->name));
1146	na->ifp = ifp;
1147	na->num_tx_desc = num_tx_desc;
1148	na->num_rx_desc = num_rx_desc;
1149	na->rx_buf_maxsize = 32768;
1150	na->nm_register = &generic_netmap_register;
1151	na->nm_txsync = &generic_netmap_txsync;
1152	na->nm_rxsync = &generic_netmap_rxsync;
1153	na->nm_dtor = &generic_netmap_dtor;
1154	/* when using generic, NAF_NETMAP_ON is set so we force
1155	 * NAF_SKIP_INTR to use the regular interrupt handler
1156	 */
1157	na->na_flags = NAF_SKIP_INTR | NAF_HOST_RINGS;
1158
1159	nm_prdis("[GNA] num_tx_queues(%d), real_num_tx_queues(%d), len(%lu)",
1160			ifp->num_tx_queues, ifp->real_num_tx_queues,
1161			ifp->tx_queue_len);
1162	nm_prdis("[GNA] num_rx_queues(%d), real_num_rx_queues(%d)",
1163			ifp->num_rx_queues, ifp->real_num_rx_queues);
1164
1165	nm_os_generic_find_num_queues(ifp, &na->num_tx_rings, &na->num_rx_rings);
1166
1167	retval = netmap_attach_common(na);
1168	if (retval) {
1169		nm_os_free(gna);
1170		return retval;
1171	}
1172
1173	if (NM_NA_VALID(ifp)) {
1174		gna->prev = NA(ifp); /* save old na */
1175		netmap_adapter_get(gna->prev);
1176	}
1177	NM_ATTACH_NA(ifp, na);
1178
1179	nm_os_generic_set_features(gna);
1180
1181	nm_prinf("Emulated adapter for %s created (prev was %s)", na->name,
1182	    gna->prev ? gna->prev->name : "NULL");
1183
1184	return retval;
1185}
1186