xref: /freebsd/sys/dev/netmap/if_ptnet.c (revision ab40f58ccfe6c07ebefddc72f4661a52fe746353)
1 /*-
2  * Copyright (c) 2016, Vincenzo Maffione
3  * 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 unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  *
26  * $FreeBSD$
27  */
28 
29 /* Driver for ptnet paravirtualized network device. */
30 
31 #include <sys/cdefs.h>
32 
33 #include <sys/types.h>
34 #include <sys/param.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
37 #include <sys/sockio.h>
38 #include <sys/mbuf.h>
39 #include <sys/malloc.h>
40 #include <sys/module.h>
41 #include <sys/socket.h>
42 #include <sys/sysctl.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/taskqueue.h>
46 #include <sys/smp.h>
47 #include <sys/time.h>
48 #include <machine/smp.h>
49 
50 #include <vm/uma.h>
51 #include <vm/vm.h>
52 #include <vm/pmap.h>
53 
54 #include <net/ethernet.h>
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_arp.h>
58 #include <net/if_dl.h>
59 #include <net/if_types.h>
60 #include <net/if_media.h>
61 #include <net/if_vlan_var.h>
62 #include <net/bpf.h>
63 
64 #include <netinet/in_systm.h>
65 #include <netinet/in.h>
66 #include <netinet/ip.h>
67 #include <netinet/ip6.h>
68 #include <netinet6/ip6_var.h>
69 #include <netinet/udp.h>
70 #include <netinet/tcp.h>
71 #include <netinet/sctp.h>
72 
73 #include <machine/bus.h>
74 #include <machine/resource.h>
75 #include <sys/bus.h>
76 #include <sys/rman.h>
77 
78 #include <dev/pci/pcivar.h>
79 #include <dev/pci/pcireg.h>
80 
81 #include "opt_inet.h"
82 #include "opt_inet6.h"
83 
84 #include <sys/selinfo.h>
85 #include <net/netmap.h>
86 #include <dev/netmap/netmap_kern.h>
87 #include <net/netmap_virt.h>
88 #include <dev/netmap/netmap_mem2.h>
89 #include <dev/virtio/network/virtio_net.h>
90 
91 #ifndef INET
92 #error "INET not defined, cannot support offloadings"
93 #endif
94 
95 #if __FreeBSD_version >= 1100000
96 static uint64_t	ptnet_get_counter(if_t, ift_counter);
97 #else
98 typedef struct ifnet *if_t;
99 #define if_getsoftc(_ifp)   (_ifp)->if_softc
100 #endif
101 
102 //#define PTNETMAP_STATS
103 //#define DEBUG
104 #ifdef DEBUG
105 #define DBG(x) x
106 #else   /* !DEBUG */
107 #define DBG(x)
108 #endif  /* !DEBUG */
109 
110 extern int ptnet_vnet_hdr; /* Tunable parameter */
111 
112 struct ptnet_softc;
113 
114 struct ptnet_queue_stats {
115 	uint64_t	packets; /* if_[io]packets */
116 	uint64_t	bytes;	 /* if_[io]bytes */
117 	uint64_t	errors;	 /* if_[io]errors */
118 	uint64_t	iqdrops; /* if_iqdrops */
119 	uint64_t	mcasts;  /* if_[io]mcasts */
120 #ifdef PTNETMAP_STATS
121 	uint64_t	intrs;
122 	uint64_t	kicks;
123 #endif /* PTNETMAP_STATS */
124 };
125 
126 struct ptnet_queue {
127 	struct ptnet_softc		*sc;
128 	struct				resource *irq;
129 	void				*cookie;
130 	int				kring_id;
131 	struct ptnet_csb_gh		*ptgh;
132 	struct ptnet_csb_hg		*pthg;
133 	unsigned int			kick;
134 	struct mtx			lock;
135 	struct buf_ring			*bufring; /* for TX queues */
136 	struct ptnet_queue_stats	stats;
137 #ifdef PTNETMAP_STATS
138 	struct ptnet_queue_stats	last_stats;
139 #endif /* PTNETMAP_STATS */
140 	struct taskqueue		*taskq;
141 	struct task			task;
142 	char				lock_name[16];
143 };
144 
145 #define PTNET_Q_LOCK(_pq)	mtx_lock(&(_pq)->lock)
146 #define PTNET_Q_TRYLOCK(_pq)	mtx_trylock(&(_pq)->lock)
147 #define PTNET_Q_UNLOCK(_pq)	mtx_unlock(&(_pq)->lock)
148 
149 struct ptnet_softc {
150 	device_t		dev;
151 	if_t			ifp;
152 	struct ifmedia		media;
153 	struct mtx		lock;
154 	char			lock_name[16];
155 	char			hwaddr[ETHER_ADDR_LEN];
156 
157 	/* Mirror of PTFEAT register. */
158 	uint32_t		ptfeatures;
159 	unsigned int		vnet_hdr_len;
160 
161 	/* PCI BARs support. */
162 	struct resource		*iomem;
163 	struct resource		*msix_mem;
164 
165 	unsigned int		num_rings;
166 	unsigned int		num_tx_rings;
167 	struct ptnet_queue	*queues;
168 	struct ptnet_queue	*rxqueues;
169 	struct ptnet_csb_gh    *csb_gh;
170 	struct ptnet_csb_hg    *csb_hg;
171 
172 	unsigned int		min_tx_space;
173 
174 	struct netmap_pt_guest_adapter *ptna;
175 
176 	struct callout		tick;
177 #ifdef PTNETMAP_STATS
178 	struct timeval		last_ts;
179 #endif /* PTNETMAP_STATS */
180 };
181 
182 #define PTNET_CORE_LOCK(_sc)	mtx_lock(&(_sc)->lock)
183 #define PTNET_CORE_UNLOCK(_sc)	mtx_unlock(&(_sc)->lock)
184 
185 static int	ptnet_probe(device_t);
186 static int	ptnet_attach(device_t);
187 static int	ptnet_detach(device_t);
188 static int	ptnet_suspend(device_t);
189 static int	ptnet_resume(device_t);
190 static int	ptnet_shutdown(device_t);
191 
192 static void	ptnet_init(void *opaque);
193 static int	ptnet_ioctl(if_t ifp, u_long cmd, caddr_t data);
194 static int	ptnet_init_locked(struct ptnet_softc *sc);
195 static int	ptnet_stop(struct ptnet_softc *sc);
196 static int	ptnet_transmit(if_t ifp, struct mbuf *m);
197 static int	ptnet_drain_transmit_queue(struct ptnet_queue *pq,
198 					   unsigned int budget,
199 					   bool may_resched);
200 static void	ptnet_qflush(if_t ifp);
201 static void	ptnet_tx_task(void *context, int pending);
202 
203 static int	ptnet_media_change(if_t ifp);
204 static void	ptnet_media_status(if_t ifp, struct ifmediareq *ifmr);
205 #ifdef PTNETMAP_STATS
206 static void	ptnet_tick(void *opaque);
207 #endif
208 
209 static int	ptnet_irqs_init(struct ptnet_softc *sc);
210 static void	ptnet_irqs_fini(struct ptnet_softc *sc);
211 
212 static uint32_t ptnet_nm_ptctl(if_t ifp, uint32_t cmd);
213 static int      ptnet_nm_config(struct netmap_adapter *na,
214 				struct nm_config_info *info);
215 static void	ptnet_update_vnet_hdr(struct ptnet_softc *sc);
216 static int	ptnet_nm_register(struct netmap_adapter *na, int onoff);
217 static int	ptnet_nm_txsync(struct netmap_kring *kring, int flags);
218 static int	ptnet_nm_rxsync(struct netmap_kring *kring, int flags);
219 static void	ptnet_nm_intr(struct netmap_adapter *na, int onoff);
220 
221 static void	ptnet_tx_intr(void *opaque);
222 static void	ptnet_rx_intr(void *opaque);
223 
224 static unsigned	ptnet_rx_discard(struct netmap_kring *kring,
225 				 unsigned int head);
226 static int	ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget,
227 			     bool may_resched);
228 static void	ptnet_rx_task(void *context, int pending);
229 
230 #ifdef DEVICE_POLLING
231 static poll_handler_t ptnet_poll;
232 #endif
233 
234 static device_method_t ptnet_methods[] = {
235 	DEVMETHOD(device_probe,			ptnet_probe),
236 	DEVMETHOD(device_attach,		ptnet_attach),
237 	DEVMETHOD(device_detach,		ptnet_detach),
238 	DEVMETHOD(device_suspend,		ptnet_suspend),
239 	DEVMETHOD(device_resume,		ptnet_resume),
240 	DEVMETHOD(device_shutdown,		ptnet_shutdown),
241 	DEVMETHOD_END
242 };
243 
244 static driver_t ptnet_driver = {
245 	"ptnet",
246 	ptnet_methods,
247 	sizeof(struct ptnet_softc)
248 };
249 
250 /* We use (SI_ORDER_MIDDLE+2) here, see DEV_MODULE_ORDERED() invocation. */
251 static devclass_t ptnet_devclass;
252 DRIVER_MODULE_ORDERED(ptnet, pci, ptnet_driver, ptnet_devclass,
253 		      NULL, NULL, SI_ORDER_MIDDLE + 2);
254 
255 static int
256 ptnet_probe(device_t dev)
257 {
258 	if (pci_get_vendor(dev) != PTNETMAP_PCI_VENDOR_ID ||
259 		pci_get_device(dev) != PTNETMAP_PCI_NETIF_ID) {
260 		return (ENXIO);
261 	}
262 
263 	device_set_desc(dev, "ptnet network adapter");
264 
265 	return (BUS_PROBE_DEFAULT);
266 }
267 
268 static inline void ptnet_kick(struct ptnet_queue *pq)
269 {
270 #ifdef PTNETMAP_STATS
271 	pq->stats.kicks ++;
272 #endif /* PTNETMAP_STATS */
273 	bus_write_4(pq->sc->iomem, pq->kick, 0);
274 }
275 
276 #define PTNET_BUF_RING_SIZE	4096
277 #define PTNET_RX_BUDGET		512
278 #define PTNET_RX_BATCH		1
279 #define PTNET_TX_BUDGET		512
280 #define PTNET_TX_BATCH		64
281 #define PTNET_HDR_SIZE		sizeof(struct virtio_net_hdr_mrg_rxbuf)
282 #define PTNET_MAX_PKT_SIZE	65536
283 
284 #define PTNET_CSUM_OFFLOAD	(CSUM_TCP | CSUM_UDP | CSUM_SCTP)
285 #define PTNET_CSUM_OFFLOAD_IPV6	(CSUM_TCP_IPV6 | CSUM_UDP_IPV6 |\
286 				 CSUM_SCTP_IPV6)
287 #define PTNET_ALL_OFFLOAD	(CSUM_TSO | PTNET_CSUM_OFFLOAD |\
288 				 PTNET_CSUM_OFFLOAD_IPV6)
289 
290 static int
291 ptnet_attach(device_t dev)
292 {
293 	uint32_t ptfeatures = 0;
294 	unsigned int num_rx_rings, num_tx_rings;
295 	struct netmap_adapter na_arg;
296 	unsigned int nifp_offset;
297 	struct ptnet_softc *sc;
298 	if_t ifp;
299 	uint32_t macreg;
300 	int err, rid;
301 	int i;
302 
303 	sc = device_get_softc(dev);
304 	sc->dev = dev;
305 
306 	/* Setup PCI resources. */
307 	pci_enable_busmaster(dev);
308 
309 	rid = PCIR_BAR(PTNETMAP_IO_PCI_BAR);
310 	sc->iomem = bus_alloc_resource_any(dev, SYS_RES_IOPORT, &rid,
311 					   RF_ACTIVE);
312 	if (sc->iomem == NULL) {
313 		device_printf(dev, "Failed to map I/O BAR\n");
314 		return (ENXIO);
315 	}
316 
317 	/* Negotiate features with the hypervisor. */
318 	if (ptnet_vnet_hdr) {
319 		ptfeatures |= PTNETMAP_F_VNET_HDR;
320 	}
321 	bus_write_4(sc->iomem, PTNET_IO_PTFEAT, ptfeatures); /* wanted */
322 	ptfeatures = bus_read_4(sc->iomem, PTNET_IO_PTFEAT); /* acked */
323 	sc->ptfeatures = ptfeatures;
324 
325 	num_tx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_RINGS);
326 	num_rx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_RINGS);
327 	sc->num_rings = num_tx_rings + num_rx_rings;
328 	sc->num_tx_rings = num_tx_rings;
329 
330 	if (sc->num_rings * sizeof(struct ptnet_csb_gh) > PAGE_SIZE) {
331 		device_printf(dev, "CSB cannot handle that many rings (%u)\n",
332 				sc->num_rings);
333 		err = ENOMEM;
334 		goto err_path;
335 	}
336 
337 	/* Allocate CSB and carry out CSB allocation protocol. */
338 	sc->csb_gh = contigmalloc(2*PAGE_SIZE, M_DEVBUF, M_NOWAIT | M_ZERO,
339 				  (size_t)0, -1UL, PAGE_SIZE, 0);
340 	if (sc->csb_gh == NULL) {
341 		device_printf(dev, "Failed to allocate CSB\n");
342 		err = ENOMEM;
343 		goto err_path;
344 	}
345 	sc->csb_hg = (struct ptnet_csb_hg *)(((char *)sc->csb_gh) + PAGE_SIZE);
346 
347 	{
348 		/*
349 		 * We use uint64_t rather than vm_paddr_t since we
350 		 * need 64 bit addresses even on 32 bit platforms.
351 		 */
352 		uint64_t paddr = vtophys(sc->csb_gh);
353 
354 		/* CSB allocation protocol: write to BAH first, then
355 		 * to BAL (for both GH and HG sections). */
356 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAH,
357 				(paddr >> 32) & 0xffffffff);
358 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAL,
359 				paddr & 0xffffffff);
360 		paddr = vtophys(sc->csb_hg);
361 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAH,
362 				(paddr >> 32) & 0xffffffff);
363 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAL,
364 				paddr & 0xffffffff);
365 	}
366 
367 	/* Allocate and initialize per-queue data structures. */
368 	sc->queues = malloc(sizeof(struct ptnet_queue) * sc->num_rings,
369 			    M_DEVBUF, M_NOWAIT | M_ZERO);
370 	if (sc->queues == NULL) {
371 		err = ENOMEM;
372 		goto err_path;
373 	}
374 	sc->rxqueues = sc->queues + num_tx_rings;
375 
376 	for (i = 0; i < sc->num_rings; i++) {
377 		struct ptnet_queue *pq = sc->queues + i;
378 
379 		pq->sc = sc;
380 		pq->kring_id = i;
381 		pq->kick = PTNET_IO_KICK_BASE + 4 * i;
382 		pq->ptgh = sc->csb_gh + i;
383 		pq->pthg = sc->csb_hg + i;
384 		snprintf(pq->lock_name, sizeof(pq->lock_name), "%s-%d",
385 			 device_get_nameunit(dev), i);
386 		mtx_init(&pq->lock, pq->lock_name, NULL, MTX_DEF);
387 		if (i >= num_tx_rings) {
388 			/* RX queue: fix kring_id. */
389 			pq->kring_id -= num_tx_rings;
390 		} else {
391 			/* TX queue: allocate buf_ring. */
392 			pq->bufring = buf_ring_alloc(PTNET_BUF_RING_SIZE,
393 						M_DEVBUF, M_NOWAIT, &pq->lock);
394 			if (pq->bufring == NULL) {
395 				err = ENOMEM;
396 				goto err_path;
397 			}
398 		}
399 	}
400 
401 	sc->min_tx_space = 64; /* Safe initial value. */
402 
403 	err = ptnet_irqs_init(sc);
404 	if (err) {
405 		goto err_path;
406 	}
407 
408 	/* Setup Ethernet interface. */
409 	sc->ifp = ifp = if_alloc(IFT_ETHER);
410 	if (ifp == NULL) {
411 		device_printf(dev, "Failed to allocate ifnet\n");
412 		err = ENOMEM;
413 		goto err_path;
414 	}
415 
416 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
417 	ifp->if_baudrate = IF_Gbps(10);
418 	ifp->if_softc = sc;
419 	ifp->if_flags = IFF_BROADCAST | IFF_MULTICAST | IFF_SIMPLEX;
420 	ifp->if_init = ptnet_init;
421 	ifp->if_ioctl = ptnet_ioctl;
422 #if __FreeBSD_version >= 1100000
423 	ifp->if_get_counter = ptnet_get_counter;
424 #endif
425 	ifp->if_transmit = ptnet_transmit;
426 	ifp->if_qflush = ptnet_qflush;
427 
428 	ifmedia_init(&sc->media, IFM_IMASK, ptnet_media_change,
429 		     ptnet_media_status);
430 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T | IFM_FDX, 0, NULL);
431 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T | IFM_FDX);
432 
433 	macreg = bus_read_4(sc->iomem, PTNET_IO_MAC_HI);
434 	sc->hwaddr[0] = (macreg >> 8) & 0xff;
435 	sc->hwaddr[1] = macreg & 0xff;
436 	macreg = bus_read_4(sc->iomem, PTNET_IO_MAC_LO);
437 	sc->hwaddr[2] = (macreg >> 24) & 0xff;
438 	sc->hwaddr[3] = (macreg >> 16) & 0xff;
439 	sc->hwaddr[4] = (macreg >> 8) & 0xff;
440 	sc->hwaddr[5] = macreg & 0xff;
441 
442 	ether_ifattach(ifp, sc->hwaddr);
443 
444 	ifp->if_hdrlen = sizeof(struct ether_vlan_header);
445 	ifp->if_capabilities |= IFCAP_JUMBO_MTU | IFCAP_VLAN_MTU;
446 
447 	if (sc->ptfeatures & PTNETMAP_F_VNET_HDR) {
448 		/* Similarly to what the vtnet driver does, we can emulate
449 		 * VLAN offloadings by inserting and removing the 802.1Q
450 		 * header during transmit and receive. We are then able
451 		 * to do checksum offloading of VLAN frames. */
452 		ifp->if_capabilities |= IFCAP_HWCSUM | IFCAP_HWCSUM_IPV6
453 					| IFCAP_VLAN_HWCSUM
454 					| IFCAP_TSO | IFCAP_LRO
455 					| IFCAP_VLAN_HWTSO
456 					| IFCAP_VLAN_HWTAGGING;
457 	}
458 
459 	ifp->if_capenable = ifp->if_capabilities;
460 #ifdef DEVICE_POLLING
461 	/* Don't enable polling by default. */
462 	ifp->if_capabilities |= IFCAP_POLLING;
463 #endif
464 	snprintf(sc->lock_name, sizeof(sc->lock_name),
465 		 "%s", device_get_nameunit(dev));
466 	mtx_init(&sc->lock, sc->lock_name, "ptnet core lock", MTX_DEF);
467 	callout_init_mtx(&sc->tick, &sc->lock, 0);
468 
469 	/* Prepare a netmap_adapter struct instance to do netmap_attach(). */
470 	nifp_offset = bus_read_4(sc->iomem, PTNET_IO_NIFP_OFS);
471 	memset(&na_arg, 0, sizeof(na_arg));
472 	na_arg.ifp = ifp;
473 	na_arg.num_tx_desc = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_SLOTS);
474 	na_arg.num_rx_desc = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_SLOTS);
475 	na_arg.num_tx_rings = num_tx_rings;
476 	na_arg.num_rx_rings = num_rx_rings;
477 	na_arg.nm_config = ptnet_nm_config;
478 	na_arg.nm_krings_create = ptnet_nm_krings_create;
479 	na_arg.nm_krings_delete = ptnet_nm_krings_delete;
480 	na_arg.nm_dtor = ptnet_nm_dtor;
481 	na_arg.nm_intr = ptnet_nm_intr;
482 	na_arg.nm_register = ptnet_nm_register;
483 	na_arg.nm_txsync = ptnet_nm_txsync;
484 	na_arg.nm_rxsync = ptnet_nm_rxsync;
485 
486 	netmap_pt_guest_attach(&na_arg, nifp_offset,
487                                 bus_read_4(sc->iomem, PTNET_IO_HOSTMEMID));
488 
489 	/* Now a netmap adapter for this ifp has been allocated, and it
490 	 * can be accessed through NA(ifp). We also have to initialize the CSB
491 	 * pointer. */
492 	sc->ptna = (struct netmap_pt_guest_adapter *)NA(ifp);
493 
494 	/* If virtio-net header was negotiated, set the virt_hdr_len field in
495 	 * the netmap adapter, to inform users that this netmap adapter requires
496 	 * the application to deal with the headers. */
497 	ptnet_update_vnet_hdr(sc);
498 
499 	device_printf(dev, "%s() completed\n", __func__);
500 
501 	return (0);
502 
503 err_path:
504 	ptnet_detach(dev);
505 	return err;
506 }
507 
508 static int
509 ptnet_detach(device_t dev)
510 {
511 	struct ptnet_softc *sc = device_get_softc(dev);
512 	int i;
513 
514 #ifdef DEVICE_POLLING
515 	if (sc->ifp->if_capenable & IFCAP_POLLING) {
516 		ether_poll_deregister(sc->ifp);
517 	}
518 #endif
519 	callout_drain(&sc->tick);
520 
521 	if (sc->queues) {
522 		/* Drain taskqueues before calling if_detach. */
523 		for (i = 0; i < sc->num_rings; i++) {
524 			struct ptnet_queue *pq = sc->queues + i;
525 
526 			if (pq->taskq) {
527 				taskqueue_drain(pq->taskq, &pq->task);
528 			}
529 		}
530 	}
531 
532 	if (sc->ifp) {
533 		ether_ifdetach(sc->ifp);
534 
535 		/* Uninitialize netmap adapters for this device. */
536 		netmap_detach(sc->ifp);
537 
538 		ifmedia_removeall(&sc->media);
539 		if_free(sc->ifp);
540 		sc->ifp = NULL;
541 	}
542 
543 	ptnet_irqs_fini(sc);
544 
545 	if (sc->csb_gh) {
546 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAH, 0);
547 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAL, 0);
548 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAH, 0);
549 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAL, 0);
550 		contigfree(sc->csb_gh, 2*PAGE_SIZE, M_DEVBUF);
551 		sc->csb_gh = NULL;
552 		sc->csb_hg = NULL;
553 	}
554 
555 	if (sc->queues) {
556 		for (i = 0; i < sc->num_rings; i++) {
557 			struct ptnet_queue *pq = sc->queues + i;
558 
559 			if (mtx_initialized(&pq->lock)) {
560 				mtx_destroy(&pq->lock);
561 			}
562 			if (pq->bufring != NULL) {
563 				buf_ring_free(pq->bufring, M_DEVBUF);
564 			}
565 		}
566 		free(sc->queues, M_DEVBUF);
567 		sc->queues = NULL;
568 	}
569 
570 	if (sc->iomem) {
571 		bus_release_resource(dev, SYS_RES_IOPORT,
572 				     PCIR_BAR(PTNETMAP_IO_PCI_BAR), sc->iomem);
573 		sc->iomem = NULL;
574 	}
575 
576 	mtx_destroy(&sc->lock);
577 
578 	device_printf(dev, "%s() completed\n", __func__);
579 
580 	return (0);
581 }
582 
583 static int
584 ptnet_suspend(device_t dev)
585 {
586 	struct ptnet_softc *sc;
587 
588 	sc = device_get_softc(dev);
589 	(void)sc;
590 
591 	return (0);
592 }
593 
594 static int
595 ptnet_resume(device_t dev)
596 {
597 	struct ptnet_softc *sc;
598 
599 	sc = device_get_softc(dev);
600 	(void)sc;
601 
602 	return (0);
603 }
604 
605 static int
606 ptnet_shutdown(device_t dev)
607 {
608 	/*
609 	 * Suspend already does all of what we need to
610 	 * do here; we just never expect to be resumed.
611 	 */
612 	return (ptnet_suspend(dev));
613 }
614 
615 static int
616 ptnet_irqs_init(struct ptnet_softc *sc)
617 {
618 	int rid = PCIR_BAR(PTNETMAP_MSIX_PCI_BAR);
619 	int nvecs = sc->num_rings;
620 	device_t dev = sc->dev;
621 	int err = ENOSPC;
622 	int cpu_cur;
623 	int i;
624 
625 	if (pci_find_cap(dev, PCIY_MSIX, NULL) != 0)  {
626 		device_printf(dev, "Could not find MSI-X capability\n");
627 		return (ENXIO);
628 	}
629 
630 	sc->msix_mem = bus_alloc_resource_any(dev, SYS_RES_MEMORY,
631 					      &rid, RF_ACTIVE);
632 	if (sc->msix_mem == NULL) {
633 		device_printf(dev, "Failed to allocate MSIX PCI BAR\n");
634 		return (ENXIO);
635 	}
636 
637 	if (pci_msix_count(dev) < nvecs) {
638 		device_printf(dev, "Not enough MSI-X vectors\n");
639 		goto err_path;
640 	}
641 
642 	err = pci_alloc_msix(dev, &nvecs);
643 	if (err) {
644 		device_printf(dev, "Failed to allocate MSI-X vectors\n");
645 		goto err_path;
646 	}
647 
648 	for (i = 0; i < nvecs; i++) {
649 		struct ptnet_queue *pq = sc->queues + i;
650 
651 		rid = i + 1;
652 		pq->irq = bus_alloc_resource_any(dev, SYS_RES_IRQ, &rid,
653 						 RF_ACTIVE);
654 		if (pq->irq == NULL) {
655 			device_printf(dev, "Failed to allocate interrupt "
656 					   "for queue #%d\n", i);
657 			err = ENOSPC;
658 			goto err_path;
659 		}
660 	}
661 
662 	cpu_cur = CPU_FIRST();
663 	for (i = 0; i < nvecs; i++) {
664 		struct ptnet_queue *pq = sc->queues + i;
665 		void (*handler)(void *) = ptnet_tx_intr;
666 
667 		if (i >= sc->num_tx_rings) {
668 			handler = ptnet_rx_intr;
669 		}
670 		err = bus_setup_intr(dev, pq->irq, INTR_TYPE_NET | INTR_MPSAFE,
671 				     NULL /* intr_filter */, handler,
672 				     pq, &pq->cookie);
673 		if (err) {
674 			device_printf(dev, "Failed to register intr handler "
675 					   "for queue #%d\n", i);
676 			goto err_path;
677 		}
678 
679 		bus_describe_intr(dev, pq->irq, pq->cookie, "q%d", i);
680 #if 0
681 		bus_bind_intr(sc->dev, pq->irq, cpu_cur);
682 #endif
683 		cpu_cur = CPU_NEXT(cpu_cur);
684 	}
685 
686 	device_printf(dev, "Allocated %d MSI-X vectors\n", nvecs);
687 
688 	cpu_cur = CPU_FIRST();
689 	for (i = 0; i < nvecs; i++) {
690 		struct ptnet_queue *pq = sc->queues + i;
691 		static void (*handler)(void *context, int pending);
692 
693 		handler = (i < sc->num_tx_rings) ? ptnet_tx_task : ptnet_rx_task;
694 
695 		TASK_INIT(&pq->task, 0, handler, pq);
696 		pq->taskq = taskqueue_create_fast("ptnet_queue", M_NOWAIT,
697 					taskqueue_thread_enqueue, &pq->taskq);
698 		taskqueue_start_threads(&pq->taskq, 1, PI_NET, "%s-pq-%d",
699 					device_get_nameunit(sc->dev), cpu_cur);
700 		cpu_cur = CPU_NEXT(cpu_cur);
701 	}
702 
703 	return 0;
704 err_path:
705 	ptnet_irqs_fini(sc);
706 	return err;
707 }
708 
709 static void
710 ptnet_irqs_fini(struct ptnet_softc *sc)
711 {
712 	device_t dev = sc->dev;
713 	int i;
714 
715 	for (i = 0; i < sc->num_rings; i++) {
716 		struct ptnet_queue *pq = sc->queues + i;
717 
718 		if (pq->taskq) {
719 			taskqueue_free(pq->taskq);
720 			pq->taskq = NULL;
721 		}
722 
723 		if (pq->cookie) {
724 			bus_teardown_intr(dev, pq->irq, pq->cookie);
725 			pq->cookie = NULL;
726 		}
727 
728 		if (pq->irq) {
729 			bus_release_resource(dev, SYS_RES_IRQ, i + 1, pq->irq);
730 			pq->irq = NULL;
731 		}
732 	}
733 
734 	if (sc->msix_mem) {
735 		pci_release_msi(dev);
736 
737 		bus_release_resource(dev, SYS_RES_MEMORY,
738 				     PCIR_BAR(PTNETMAP_MSIX_PCI_BAR),
739 				     sc->msix_mem);
740 		sc->msix_mem = NULL;
741 	}
742 }
743 
744 static void
745 ptnet_init(void *opaque)
746 {
747 	struct ptnet_softc *sc = opaque;
748 
749 	PTNET_CORE_LOCK(sc);
750 	ptnet_init_locked(sc);
751 	PTNET_CORE_UNLOCK(sc);
752 }
753 
754 static int
755 ptnet_ioctl(if_t ifp, u_long cmd, caddr_t data)
756 {
757 	struct ptnet_softc *sc = if_getsoftc(ifp);
758 	device_t dev = sc->dev;
759 	struct ifreq *ifr = (struct ifreq *)data;
760 	int mask __unused, err = 0;
761 
762 	switch (cmd) {
763 	case SIOCSIFFLAGS:
764 		device_printf(dev, "SIOCSIFFLAGS %x\n", ifp->if_flags);
765 		PTNET_CORE_LOCK(sc);
766 		if (ifp->if_flags & IFF_UP) {
767 			/* Network stack wants the iff to be up. */
768 			err = ptnet_init_locked(sc);
769 		} else {
770 			/* Network stack wants the iff to be down. */
771 			err = ptnet_stop(sc);
772 		}
773 		/* We don't need to do nothing to support IFF_PROMISC,
774 		 * since that is managed by the backend port. */
775 		PTNET_CORE_UNLOCK(sc);
776 		break;
777 
778 	case SIOCSIFCAP:
779 		device_printf(dev, "SIOCSIFCAP %x %x\n",
780 			      ifr->ifr_reqcap, ifp->if_capenable);
781 		mask = ifr->ifr_reqcap ^ ifp->if_capenable;
782 #ifdef DEVICE_POLLING
783 		if (mask & IFCAP_POLLING) {
784 			struct ptnet_queue *pq;
785 			int i;
786 
787 			if (ifr->ifr_reqcap & IFCAP_POLLING) {
788 				err = ether_poll_register(ptnet_poll, ifp);
789 				if (err) {
790 					break;
791 				}
792 				/* Stop queues and sync with taskqueues. */
793 				ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
794 				for (i = 0; i < sc->num_rings; i++) {
795 					pq = sc-> queues + i;
796 					/* Make sure the worker sees the
797 					 * IFF_DRV_RUNNING down. */
798 					PTNET_Q_LOCK(pq);
799 					pq->ptgh->guest_need_kick = 0;
800 					PTNET_Q_UNLOCK(pq);
801 					/* Wait for rescheduling to finish. */
802 					if (pq->taskq) {
803 						taskqueue_drain(pq->taskq,
804 								&pq->task);
805 					}
806 				}
807 				ifp->if_drv_flags |= IFF_DRV_RUNNING;
808 			} else {
809 				err = ether_poll_deregister(ifp);
810 				for (i = 0; i < sc->num_rings; i++) {
811 					pq = sc-> queues + i;
812 					PTNET_Q_LOCK(pq);
813 					pq->ptgh->guest_need_kick = 1;
814 					PTNET_Q_UNLOCK(pq);
815 				}
816 			}
817 		}
818 #endif  /* DEVICE_POLLING */
819 		ifp->if_capenable = ifr->ifr_reqcap;
820 		break;
821 
822 	case SIOCSIFMTU:
823 		/* We support any reasonable MTU. */
824 		if (ifr->ifr_mtu < ETHERMIN ||
825 				ifr->ifr_mtu > PTNET_MAX_PKT_SIZE) {
826 			err = EINVAL;
827 		} else {
828 			PTNET_CORE_LOCK(sc);
829 			ifp->if_mtu = ifr->ifr_mtu;
830 			PTNET_CORE_UNLOCK(sc);
831 		}
832 		break;
833 
834 	case SIOCSIFMEDIA:
835 	case SIOCGIFMEDIA:
836 		err = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
837 		break;
838 
839 	default:
840 		err = ether_ioctl(ifp, cmd, data);
841 		break;
842 	}
843 
844 	return err;
845 }
846 
847 static int
848 ptnet_init_locked(struct ptnet_softc *sc)
849 {
850 	if_t ifp = sc->ifp;
851 	struct netmap_adapter *na_dr = &sc->ptna->dr.up;
852 	struct netmap_adapter *na_nm = &sc->ptna->hwup.up;
853 	unsigned int nm_buf_size;
854 	int ret;
855 
856 	if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
857 		return 0; /* nothing to do */
858 	}
859 
860 	device_printf(sc->dev, "%s\n", __func__);
861 
862 	/* Translate offload capabilities according to if_capenable. */
863 	ifp->if_hwassist = 0;
864 	if (ifp->if_capenable & IFCAP_TXCSUM)
865 		ifp->if_hwassist |= PTNET_CSUM_OFFLOAD;
866 	if (ifp->if_capenable & IFCAP_TXCSUM_IPV6)
867 		ifp->if_hwassist |= PTNET_CSUM_OFFLOAD_IPV6;
868 	if (ifp->if_capenable & IFCAP_TSO4)
869 		ifp->if_hwassist |= CSUM_IP_TSO;
870 	if (ifp->if_capenable & IFCAP_TSO6)
871 		ifp->if_hwassist |= CSUM_IP6_TSO;
872 
873 	/*
874 	 * Prepare the interface for netmap mode access.
875 	 */
876 	netmap_update_config(na_dr);
877 
878 	ret = netmap_mem_finalize(na_dr->nm_mem, na_dr);
879 	if (ret) {
880 		device_printf(sc->dev, "netmap_mem_finalize() failed\n");
881 		return ret;
882 	}
883 
884 	if (sc->ptna->backend_regifs == 0) {
885 		ret = ptnet_nm_krings_create(na_nm);
886 		if (ret) {
887 			device_printf(sc->dev, "ptnet_nm_krings_create() "
888 					       "failed\n");
889 			goto err_mem_finalize;
890 		}
891 
892 		ret = netmap_mem_rings_create(na_dr);
893 		if (ret) {
894 			device_printf(sc->dev, "netmap_mem_rings_create() "
895 					       "failed\n");
896 			goto err_rings_create;
897 		}
898 
899 		ret = netmap_mem_get_lut(na_dr->nm_mem, &na_dr->na_lut);
900 		if (ret) {
901 			device_printf(sc->dev, "netmap_mem_get_lut() "
902 					       "failed\n");
903 			goto err_get_lut;
904 		}
905 	}
906 
907 	ret = ptnet_nm_register(na_dr, 1 /* on */);
908 	if (ret) {
909 		goto err_register;
910 	}
911 
912 	nm_buf_size = NETMAP_BUF_SIZE(na_dr);
913 
914 	KASSERT(nm_buf_size > 0, ("Invalid netmap buffer size"));
915 	sc->min_tx_space = PTNET_MAX_PKT_SIZE / nm_buf_size + 2;
916 	device_printf(sc->dev, "%s: min_tx_space = %u\n", __func__,
917 		      sc->min_tx_space);
918 #ifdef PTNETMAP_STATS
919 	callout_reset(&sc->tick, hz, ptnet_tick, sc);
920 #endif
921 
922 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
923 
924 	return 0;
925 
926 err_register:
927 	memset(&na_dr->na_lut, 0, sizeof(na_dr->na_lut));
928 err_get_lut:
929 	netmap_mem_rings_delete(na_dr);
930 err_rings_create:
931 	ptnet_nm_krings_delete(na_nm);
932 err_mem_finalize:
933 	netmap_mem_deref(na_dr->nm_mem, na_dr);
934 
935 	return ret;
936 }
937 
938 /* To be called under core lock. */
939 static int
940 ptnet_stop(struct ptnet_softc *sc)
941 {
942 	if_t ifp = sc->ifp;
943 	struct netmap_adapter *na_dr = &sc->ptna->dr.up;
944 	struct netmap_adapter *na_nm = &sc->ptna->hwup.up;
945 	int i;
946 
947 	device_printf(sc->dev, "%s\n", __func__);
948 
949 	if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) {
950 		return 0; /* nothing to do */
951 	}
952 
953 	/* Clear the driver-ready flag, and synchronize with all the queues,
954 	 * so that after this loop we are sure nobody is working anymore with
955 	 * the device. This scheme is taken from the vtnet driver. */
956 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
957 	callout_stop(&sc->tick);
958 	for (i = 0; i < sc->num_rings; i++) {
959 		PTNET_Q_LOCK(sc->queues + i);
960 		PTNET_Q_UNLOCK(sc->queues + i);
961 	}
962 
963 	ptnet_nm_register(na_dr, 0 /* off */);
964 
965 	if (sc->ptna->backend_regifs == 0) {
966 		netmap_mem_rings_delete(na_dr);
967 		ptnet_nm_krings_delete(na_nm);
968 	}
969 	netmap_mem_deref(na_dr->nm_mem, na_dr);
970 
971 	return 0;
972 }
973 
974 static void
975 ptnet_qflush(if_t ifp)
976 {
977 	struct ptnet_softc *sc = if_getsoftc(ifp);
978 	int i;
979 
980 	/* Flush all the bufrings and do the interface flush. */
981 	for (i = 0; i < sc->num_rings; i++) {
982 		struct ptnet_queue *pq = sc->queues + i;
983 		struct mbuf *m;
984 
985 		PTNET_Q_LOCK(pq);
986 		if (pq->bufring) {
987 			while ((m = buf_ring_dequeue_sc(pq->bufring))) {
988 				m_freem(m);
989 			}
990 		}
991 		PTNET_Q_UNLOCK(pq);
992 	}
993 
994 	if_qflush(ifp);
995 }
996 
997 static int
998 ptnet_media_change(if_t ifp)
999 {
1000 	struct ptnet_softc *sc = if_getsoftc(ifp);
1001 	struct ifmedia *ifm = &sc->media;
1002 
1003 	if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER) {
1004 		return EINVAL;
1005 	}
1006 
1007 	return 0;
1008 }
1009 
1010 #if __FreeBSD_version >= 1100000
1011 static uint64_t
1012 ptnet_get_counter(if_t ifp, ift_counter cnt)
1013 {
1014 	struct ptnet_softc *sc = if_getsoftc(ifp);
1015 	struct ptnet_queue_stats stats[2];
1016 	int i;
1017 
1018 	/* Accumulate statistics over the queues. */
1019 	memset(stats, 0, sizeof(stats));
1020 	for (i = 0; i < sc->num_rings; i++) {
1021 		struct ptnet_queue *pq = sc->queues + i;
1022 		int idx = (i < sc->num_tx_rings) ? 0 : 1;
1023 
1024 		stats[idx].packets	+= pq->stats.packets;
1025 		stats[idx].bytes	+= pq->stats.bytes;
1026 		stats[idx].errors	+= pq->stats.errors;
1027 		stats[idx].iqdrops	+= pq->stats.iqdrops;
1028 		stats[idx].mcasts	+= pq->stats.mcasts;
1029 	}
1030 
1031 	switch (cnt) {
1032 	case IFCOUNTER_IPACKETS:
1033 		return (stats[1].packets);
1034 	case IFCOUNTER_IQDROPS:
1035 		return (stats[1].iqdrops);
1036 	case IFCOUNTER_IERRORS:
1037 		return (stats[1].errors);
1038 	case IFCOUNTER_OPACKETS:
1039 		return (stats[0].packets);
1040 	case IFCOUNTER_OBYTES:
1041 		return (stats[0].bytes);
1042 	case IFCOUNTER_OMCASTS:
1043 		return (stats[0].mcasts);
1044 	default:
1045 		return (if_get_counter_default(ifp, cnt));
1046 	}
1047 }
1048 #endif
1049 
1050 
1051 #ifdef PTNETMAP_STATS
1052 /* Called under core lock. */
1053 static void
1054 ptnet_tick(void *opaque)
1055 {
1056 	struct ptnet_softc *sc = opaque;
1057 	int i;
1058 
1059 	for (i = 0; i < sc->num_rings; i++) {
1060 		struct ptnet_queue *pq = sc->queues + i;
1061 		struct ptnet_queue_stats cur = pq->stats;
1062 		struct timeval now;
1063 		unsigned int delta;
1064 
1065 		microtime(&now);
1066 		delta = now.tv_usec - sc->last_ts.tv_usec +
1067 			(now.tv_sec - sc->last_ts.tv_sec) * 1000000;
1068 		delta /= 1000; /* in milliseconds */
1069 
1070 		if (delta == 0)
1071 			continue;
1072 
1073 		device_printf(sc->dev, "#%d[%u ms]:pkts %lu, kicks %lu, "
1074 			      "intr %lu\n", i, delta,
1075 			      (cur.packets - pq->last_stats.packets),
1076 			      (cur.kicks - pq->last_stats.kicks),
1077 			      (cur.intrs - pq->last_stats.intrs));
1078 		pq->last_stats = cur;
1079 	}
1080 	microtime(&sc->last_ts);
1081 	callout_schedule(&sc->tick, hz);
1082 }
1083 #endif /* PTNETMAP_STATS */
1084 
1085 static void
1086 ptnet_media_status(if_t ifp, struct ifmediareq *ifmr)
1087 {
1088 	/* We are always active, as the backend netmap port is
1089 	 * always open in netmap mode. */
1090 	ifmr->ifm_status = IFM_AVALID | IFM_ACTIVE;
1091 	ifmr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
1092 }
1093 
1094 static uint32_t
1095 ptnet_nm_ptctl(if_t ifp, uint32_t cmd)
1096 {
1097 	struct ptnet_softc *sc = if_getsoftc(ifp);
1098 	/*
1099 	 * Write a command and read back error status,
1100 	 * with zero meaning success.
1101 	 */
1102 	bus_write_4(sc->iomem, PTNET_IO_PTCTL, cmd);
1103 	return bus_read_4(sc->iomem, PTNET_IO_PTCTL);
1104 }
1105 
1106 static int
1107 ptnet_nm_config(struct netmap_adapter *na, struct nm_config_info *info)
1108 {
1109 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1110 
1111 	info->num_tx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_RINGS);
1112 	info->num_rx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_RINGS);
1113 	info->num_tx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_SLOTS);
1114 	info->num_rx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_SLOTS);
1115 	info->rx_buf_maxsize = NETMAP_BUF_SIZE(na);
1116 
1117 	device_printf(sc->dev, "txr %u, rxr %u, txd %u, rxd %u, rxbufsz %u\n",
1118 			info->num_tx_rings, info->num_rx_rings,
1119 			info->num_tx_descs, info->num_rx_descs,
1120 			info->rx_buf_maxsize);
1121 
1122 	return 0;
1123 }
1124 
1125 static void
1126 ptnet_sync_from_csb(struct ptnet_softc *sc, struct netmap_adapter *na)
1127 {
1128 	int i;
1129 
1130 	/* Sync krings from the host, reading from
1131 	 * CSB. */
1132 	for (i = 0; i < sc->num_rings; i++) {
1133 		struct ptnet_csb_gh *ptgh = sc->queues[i].ptgh;
1134 		struct ptnet_csb_hg *pthg = sc->queues[i].pthg;
1135 		struct netmap_kring *kring;
1136 
1137 		if (i < na->num_tx_rings) {
1138 			kring = na->tx_rings[i];
1139 		} else {
1140 			kring = na->rx_rings[i - na->num_tx_rings];
1141 		}
1142 		kring->rhead = kring->ring->head = ptgh->head;
1143 		kring->rcur = kring->ring->cur = ptgh->cur;
1144 		kring->nr_hwcur = pthg->hwcur;
1145 		kring->nr_hwtail = kring->rtail =
1146 			kring->ring->tail = pthg->hwtail;
1147 
1148 		ND("%d,%d: csb {hc %u h %u c %u ht %u}", t, i,
1149 		   pthg->hwcur, ptgh->head, ptgh->cur,
1150 		   pthg->hwtail);
1151 		ND("%d,%d: kring {hc %u rh %u rc %u h %u c %u ht %u rt %u t %u}",
1152 		   t, i, kring->nr_hwcur, kring->rhead, kring->rcur,
1153 		   kring->ring->head, kring->ring->cur, kring->nr_hwtail,
1154 		   kring->rtail, kring->ring->tail);
1155 	}
1156 }
1157 
1158 static void
1159 ptnet_update_vnet_hdr(struct ptnet_softc *sc)
1160 {
1161 	unsigned int wanted_hdr_len = ptnet_vnet_hdr ? PTNET_HDR_SIZE : 0;
1162 
1163 	bus_write_4(sc->iomem, PTNET_IO_VNET_HDR_LEN, wanted_hdr_len);
1164 	sc->vnet_hdr_len = bus_read_4(sc->iomem, PTNET_IO_VNET_HDR_LEN);
1165 	sc->ptna->hwup.up.virt_hdr_len = sc->vnet_hdr_len;
1166 }
1167 
1168 static int
1169 ptnet_nm_register(struct netmap_adapter *na, int onoff)
1170 {
1171 	/* device-specific */
1172 	if_t ifp = na->ifp;
1173 	struct ptnet_softc *sc = if_getsoftc(ifp);
1174 	int native = (na == &sc->ptna->hwup.up);
1175 	struct ptnet_queue *pq;
1176 	enum txrx t;
1177 	int ret = 0;
1178 	int i;
1179 
1180 	if (!onoff) {
1181 		sc->ptna->backend_regifs--;
1182 	}
1183 
1184 	/* If this is the last netmap client, guest interrupt enable flags may
1185 	 * be in arbitrary state. Since these flags are going to be used also
1186 	 * by the netdevice driver, we have to make sure to start with
1187 	 * notifications enabled. Also, schedule NAPI to flush pending packets
1188 	 * in the RX rings, since we will not receive further interrupts
1189 	 * until these will be processed. */
1190 	if (native && !onoff && na->active_fds == 0) {
1191 		D("Exit netmap mode, re-enable interrupts");
1192 		for (i = 0; i < sc->num_rings; i++) {
1193 			pq = sc->queues + i;
1194 			pq->ptgh->guest_need_kick = 1;
1195 		}
1196 	}
1197 
1198 	if (onoff) {
1199 		if (sc->ptna->backend_regifs == 0) {
1200 			/* Initialize notification enable fields in the CSB. */
1201 			for (i = 0; i < sc->num_rings; i++) {
1202 				pq = sc->queues + i;
1203 				pq->pthg->host_need_kick = 1;
1204 				pq->ptgh->guest_need_kick =
1205 					(!(ifp->if_capenable & IFCAP_POLLING)
1206 						&& i >= sc->num_tx_rings);
1207 			}
1208 
1209 			/* Set the virtio-net header length. */
1210 			ptnet_update_vnet_hdr(sc);
1211 
1212 			/* Make sure the host adapter passed through is ready
1213 			 * for txsync/rxsync. */
1214 			ret = ptnet_nm_ptctl(ifp, PTNETMAP_PTCTL_CREATE);
1215 			if (ret) {
1216 				return ret;
1217 			}
1218 		}
1219 
1220 		/* Sync from CSB must be done after REGIF PTCTL. Skip this
1221 		 * step only if this is a netmap client and it is not the
1222 		 * first one. */
1223 		if ((!native && sc->ptna->backend_regifs == 0) ||
1224 				(native && na->active_fds == 0)) {
1225 			ptnet_sync_from_csb(sc, na);
1226 		}
1227 
1228 		/* If not native, don't call nm_set_native_flags, since we don't want
1229 		 * to replace if_transmit method, nor set NAF_NETMAP_ON */
1230 		if (native) {
1231 			for_rx_tx(t) {
1232 				for (i = 0; i <= nma_get_nrings(na, t); i++) {
1233 					struct netmap_kring *kring = NMR(na, t)[i];
1234 
1235 					if (nm_kring_pending_on(kring)) {
1236 						kring->nr_mode = NKR_NETMAP_ON;
1237 					}
1238 				}
1239 			}
1240 			nm_set_native_flags(na);
1241 		}
1242 
1243 	} else {
1244 		if (native) {
1245 			nm_clear_native_flags(na);
1246 			for_rx_tx(t) {
1247 				for (i = 0; i <= nma_get_nrings(na, t); i++) {
1248 					struct netmap_kring *kring = NMR(na, t)[i];
1249 
1250 					if (nm_kring_pending_off(kring)) {
1251 						kring->nr_mode = NKR_NETMAP_OFF;
1252 					}
1253 				}
1254 			}
1255 		}
1256 
1257 		/* Sync from CSB must be done before UNREGIF PTCTL, on the last
1258 		 * netmap client. */
1259 		if (native && na->active_fds == 0) {
1260 			ptnet_sync_from_csb(sc, na);
1261 		}
1262 
1263 		if (sc->ptna->backend_regifs == 0) {
1264 			ret = ptnet_nm_ptctl(ifp, PTNETMAP_PTCTL_DELETE);
1265 		}
1266 	}
1267 
1268 	if (onoff) {
1269 		sc->ptna->backend_regifs++;
1270 	}
1271 
1272 	return ret;
1273 }
1274 
1275 static int
1276 ptnet_nm_txsync(struct netmap_kring *kring, int flags)
1277 {
1278 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1279 	struct ptnet_queue *pq = sc->queues + kring->ring_id;
1280 	bool notify;
1281 
1282 	notify = netmap_pt_guest_txsync(pq->ptgh, pq->pthg, kring, flags);
1283 	if (notify) {
1284 		ptnet_kick(pq);
1285 	}
1286 
1287 	return 0;
1288 }
1289 
1290 static int
1291 ptnet_nm_rxsync(struct netmap_kring *kring, int flags)
1292 {
1293 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1294 	struct ptnet_queue *pq = sc->rxqueues + kring->ring_id;
1295 	bool notify;
1296 
1297 	notify = netmap_pt_guest_rxsync(pq->ptgh, pq->pthg, kring, flags);
1298 	if (notify) {
1299 		ptnet_kick(pq);
1300 	}
1301 
1302 	return 0;
1303 }
1304 
1305 static void
1306 ptnet_nm_intr(struct netmap_adapter *na, int onoff)
1307 {
1308 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1309 	int i;
1310 
1311 	for (i = 0; i < sc->num_rings; i++) {
1312 		struct ptnet_queue *pq = sc->queues + i;
1313 		pq->ptgh->guest_need_kick = onoff;
1314 	}
1315 }
1316 
1317 static void
1318 ptnet_tx_intr(void *opaque)
1319 {
1320 	struct ptnet_queue *pq = opaque;
1321 	struct ptnet_softc *sc = pq->sc;
1322 
1323 	DBG(device_printf(sc->dev, "Tx interrupt #%d\n", pq->kring_id));
1324 #ifdef PTNETMAP_STATS
1325 	pq->stats.intrs ++;
1326 #endif /* PTNETMAP_STATS */
1327 
1328 	if (netmap_tx_irq(sc->ifp, pq->kring_id) != NM_IRQ_PASS) {
1329 		return;
1330 	}
1331 
1332 	/* Schedule the tasqueue to flush process transmissions requests.
1333 	 * However, vtnet, if_em and if_igb just call ptnet_transmit() here,
1334 	 * at least when using MSI-X interrupts. The if_em driver, instead
1335 	 * schedule taskqueue when using legacy interrupts. */
1336 	taskqueue_enqueue(pq->taskq, &pq->task);
1337 }
1338 
1339 static void
1340 ptnet_rx_intr(void *opaque)
1341 {
1342 	struct ptnet_queue *pq = opaque;
1343 	struct ptnet_softc *sc = pq->sc;
1344 	unsigned int unused;
1345 
1346 	DBG(device_printf(sc->dev, "Rx interrupt #%d\n", pq->kring_id));
1347 #ifdef PTNETMAP_STATS
1348 	pq->stats.intrs ++;
1349 #endif /* PTNETMAP_STATS */
1350 
1351 	if (netmap_rx_irq(sc->ifp, pq->kring_id, &unused) != NM_IRQ_PASS) {
1352 		return;
1353 	}
1354 
1355 	/* Like vtnet, if_igb and if_em drivers when using MSI-X interrupts,
1356 	 * receive-side processing is executed directly in the interrupt
1357 	 * service routine. Alternatively, we may schedule the taskqueue. */
1358 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
1359 }
1360 
1361 /* The following offloadings-related functions are taken from the vtnet
1362  * driver, but the same functionality is required for the ptnet driver.
1363  * As a temporary solution, I copied this code from vtnet and I started
1364  * to generalize it (taking away driver-specific statistic accounting),
1365  * making as little modifications as possible.
1366  * In the future we need to share these functions between vtnet and ptnet.
1367  */
1368 static int
1369 ptnet_tx_offload_ctx(struct mbuf *m, int *etype, int *proto, int *start)
1370 {
1371 	struct ether_vlan_header *evh;
1372 	int offset;
1373 
1374 	evh = mtod(m, struct ether_vlan_header *);
1375 	if (evh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
1376 		/* BMV: We should handle nested VLAN tags too. */
1377 		*etype = ntohs(evh->evl_proto);
1378 		offset = sizeof(struct ether_vlan_header);
1379 	} else {
1380 		*etype = ntohs(evh->evl_encap_proto);
1381 		offset = sizeof(struct ether_header);
1382 	}
1383 
1384 	switch (*etype) {
1385 #if defined(INET)
1386 	case ETHERTYPE_IP: {
1387 		struct ip *ip, iphdr;
1388 		if (__predict_false(m->m_len < offset + sizeof(struct ip))) {
1389 			m_copydata(m, offset, sizeof(struct ip),
1390 			    (caddr_t) &iphdr);
1391 			ip = &iphdr;
1392 		} else
1393 			ip = (struct ip *)(m->m_data + offset);
1394 		*proto = ip->ip_p;
1395 		*start = offset + (ip->ip_hl << 2);
1396 		break;
1397 	}
1398 #endif
1399 #if defined(INET6)
1400 	case ETHERTYPE_IPV6:
1401 		*proto = -1;
1402 		*start = ip6_lasthdr(m, offset, IPPROTO_IPV6, proto);
1403 		/* Assert the network stack sent us a valid packet. */
1404 		KASSERT(*start > offset,
1405 		    ("%s: mbuf %p start %d offset %d proto %d", __func__, m,
1406 		    *start, offset, *proto));
1407 		break;
1408 #endif
1409 	default:
1410 		/* Here we should increment the tx_csum_bad_ethtype counter. */
1411 		return (EINVAL);
1412 	}
1413 
1414 	return (0);
1415 }
1416 
1417 static int
1418 ptnet_tx_offload_tso(if_t ifp, struct mbuf *m, int eth_type,
1419 		     int offset, bool allow_ecn, struct virtio_net_hdr *hdr)
1420 {
1421 	static struct timeval lastecn;
1422 	static int curecn;
1423 	struct tcphdr *tcp, tcphdr;
1424 
1425 	if (__predict_false(m->m_len < offset + sizeof(struct tcphdr))) {
1426 		m_copydata(m, offset, sizeof(struct tcphdr), (caddr_t) &tcphdr);
1427 		tcp = &tcphdr;
1428 	} else
1429 		tcp = (struct tcphdr *)(m->m_data + offset);
1430 
1431 	hdr->hdr_len = offset + (tcp->th_off << 2);
1432 	hdr->gso_size = m->m_pkthdr.tso_segsz;
1433 	hdr->gso_type = eth_type == ETHERTYPE_IP ? VIRTIO_NET_HDR_GSO_TCPV4 :
1434 	    VIRTIO_NET_HDR_GSO_TCPV6;
1435 
1436 	if (tcp->th_flags & TH_CWR) {
1437 		/*
1438 		 * Drop if VIRTIO_NET_F_HOST_ECN was not negotiated. In FreeBSD,
1439 		 * ECN support is not on a per-interface basis, but globally via
1440 		 * the net.inet.tcp.ecn.enable sysctl knob. The default is off.
1441 		 */
1442 		if (!allow_ecn) {
1443 			if (ppsratecheck(&lastecn, &curecn, 1))
1444 				if_printf(ifp,
1445 				    "TSO with ECN not negotiated with host\n");
1446 			return (ENOTSUP);
1447 		}
1448 		hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
1449 	}
1450 
1451 	/* Here we should increment tx_tso counter. */
1452 
1453 	return (0);
1454 }
1455 
1456 static struct mbuf *
1457 ptnet_tx_offload(if_t ifp, struct mbuf *m, bool allow_ecn,
1458 		 struct virtio_net_hdr *hdr)
1459 {
1460 	int flags, etype, csum_start, proto, error;
1461 
1462 	flags = m->m_pkthdr.csum_flags;
1463 
1464 	error = ptnet_tx_offload_ctx(m, &etype, &proto, &csum_start);
1465 	if (error)
1466 		goto drop;
1467 
1468 	if ((etype == ETHERTYPE_IP && flags & PTNET_CSUM_OFFLOAD) ||
1469 	    (etype == ETHERTYPE_IPV6 && flags & PTNET_CSUM_OFFLOAD_IPV6)) {
1470 		/*
1471 		 * We could compare the IP protocol vs the CSUM_ flag too,
1472 		 * but that really should not be necessary.
1473 		 */
1474 		hdr->flags |= VIRTIO_NET_HDR_F_NEEDS_CSUM;
1475 		hdr->csum_start = csum_start;
1476 		hdr->csum_offset = m->m_pkthdr.csum_data;
1477 		/* Here we should increment the tx_csum counter. */
1478 	}
1479 
1480 	if (flags & CSUM_TSO) {
1481 		if (__predict_false(proto != IPPROTO_TCP)) {
1482 			/* Likely failed to correctly parse the mbuf.
1483 			 * Here we should increment the tx_tso_not_tcp
1484 			 * counter. */
1485 			goto drop;
1486 		}
1487 
1488 		KASSERT(hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM,
1489 		    ("%s: mbuf %p TSO without checksum offload %#x",
1490 		    __func__, m, flags));
1491 
1492 		error = ptnet_tx_offload_tso(ifp, m, etype, csum_start,
1493 					     allow_ecn, hdr);
1494 		if (error)
1495 			goto drop;
1496 	}
1497 
1498 	return (m);
1499 
1500 drop:
1501 	m_freem(m);
1502 	return (NULL);
1503 }
1504 
1505 static void
1506 ptnet_vlan_tag_remove(struct mbuf *m)
1507 {
1508 	struct ether_vlan_header *evh;
1509 
1510 	evh = mtod(m, struct ether_vlan_header *);
1511 	m->m_pkthdr.ether_vtag = ntohs(evh->evl_tag);
1512 	m->m_flags |= M_VLANTAG;
1513 
1514 	/* Strip the 802.1Q header. */
1515 	bcopy((char *) evh, (char *) evh + ETHER_VLAN_ENCAP_LEN,
1516 	    ETHER_HDR_LEN - ETHER_TYPE_LEN);
1517 	m_adj(m, ETHER_VLAN_ENCAP_LEN);
1518 }
1519 
1520 /*
1521  * Use the checksum offset in the VirtIO header to set the
1522  * correct CSUM_* flags.
1523  */
1524 static int
1525 ptnet_rx_csum_by_offset(struct mbuf *m, uint16_t eth_type, int ip_start,
1526 			struct virtio_net_hdr *hdr)
1527 {
1528 #if defined(INET) || defined(INET6)
1529 	int offset = hdr->csum_start + hdr->csum_offset;
1530 #endif
1531 
1532 	/* Only do a basic sanity check on the offset. */
1533 	switch (eth_type) {
1534 #if defined(INET)
1535 	case ETHERTYPE_IP:
1536 		if (__predict_false(offset < ip_start + sizeof(struct ip)))
1537 			return (1);
1538 		break;
1539 #endif
1540 #if defined(INET6)
1541 	case ETHERTYPE_IPV6:
1542 		if (__predict_false(offset < ip_start + sizeof(struct ip6_hdr)))
1543 			return (1);
1544 		break;
1545 #endif
1546 	default:
1547 		/* Here we should increment the rx_csum_bad_ethtype counter. */
1548 		return (1);
1549 	}
1550 
1551 	/*
1552 	 * Use the offset to determine the appropriate CSUM_* flags. This is
1553 	 * a bit dirty, but we can get by with it since the checksum offsets
1554 	 * happen to be different. We assume the host host does not do IPv4
1555 	 * header checksum offloading.
1556 	 */
1557 	switch (hdr->csum_offset) {
1558 	case offsetof(struct udphdr, uh_sum):
1559 	case offsetof(struct tcphdr, th_sum):
1560 		m->m_pkthdr.csum_flags |= CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
1561 		m->m_pkthdr.csum_data = 0xFFFF;
1562 		break;
1563 	case offsetof(struct sctphdr, checksum):
1564 		m->m_pkthdr.csum_flags |= CSUM_SCTP_VALID;
1565 		break;
1566 	default:
1567 		/* Here we should increment the rx_csum_bad_offset counter. */
1568 		return (1);
1569 	}
1570 
1571 	return (0);
1572 }
1573 
1574 static int
1575 ptnet_rx_csum_by_parse(struct mbuf *m, uint16_t eth_type, int ip_start,
1576 		       struct virtio_net_hdr *hdr)
1577 {
1578 	int offset, proto;
1579 
1580 	switch (eth_type) {
1581 #if defined(INET)
1582 	case ETHERTYPE_IP: {
1583 		struct ip *ip;
1584 		if (__predict_false(m->m_len < ip_start + sizeof(struct ip)))
1585 			return (1);
1586 		ip = (struct ip *)(m->m_data + ip_start);
1587 		proto = ip->ip_p;
1588 		offset = ip_start + (ip->ip_hl << 2);
1589 		break;
1590 	}
1591 #endif
1592 #if defined(INET6)
1593 	case ETHERTYPE_IPV6:
1594 		if (__predict_false(m->m_len < ip_start +
1595 		    sizeof(struct ip6_hdr)))
1596 			return (1);
1597 		offset = ip6_lasthdr(m, ip_start, IPPROTO_IPV6, &proto);
1598 		if (__predict_false(offset < 0))
1599 			return (1);
1600 		break;
1601 #endif
1602 	default:
1603 		/* Here we should increment the rx_csum_bad_ethtype counter. */
1604 		return (1);
1605 	}
1606 
1607 	switch (proto) {
1608 	case IPPROTO_TCP:
1609 		if (__predict_false(m->m_len < offset + sizeof(struct tcphdr)))
1610 			return (1);
1611 		m->m_pkthdr.csum_flags |= CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
1612 		m->m_pkthdr.csum_data = 0xFFFF;
1613 		break;
1614 	case IPPROTO_UDP:
1615 		if (__predict_false(m->m_len < offset + sizeof(struct udphdr)))
1616 			return (1);
1617 		m->m_pkthdr.csum_flags |= CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
1618 		m->m_pkthdr.csum_data = 0xFFFF;
1619 		break;
1620 	case IPPROTO_SCTP:
1621 		if (__predict_false(m->m_len < offset + sizeof(struct sctphdr)))
1622 			return (1);
1623 		m->m_pkthdr.csum_flags |= CSUM_SCTP_VALID;
1624 		break;
1625 	default:
1626 		/*
1627 		 * For the remaining protocols, FreeBSD does not support
1628 		 * checksum offloading, so the checksum will be recomputed.
1629 		 */
1630 #if 0
1631 		if_printf(ifp, "cksum offload of unsupported "
1632 		    "protocol eth_type=%#x proto=%d csum_start=%d "
1633 		    "csum_offset=%d\n", __func__, eth_type, proto,
1634 		    hdr->csum_start, hdr->csum_offset);
1635 #endif
1636 		break;
1637 	}
1638 
1639 	return (0);
1640 }
1641 
1642 /*
1643  * Set the appropriate CSUM_* flags. Unfortunately, the information
1644  * provided is not directly useful to us. The VirtIO header gives the
1645  * offset of the checksum, which is all Linux needs, but this is not
1646  * how FreeBSD does things. We are forced to peek inside the packet
1647  * a bit.
1648  *
1649  * It would be nice if VirtIO gave us the L4 protocol or if FreeBSD
1650  * could accept the offsets and let the stack figure it out.
1651  */
1652 static int
1653 ptnet_rx_csum(struct mbuf *m, struct virtio_net_hdr *hdr)
1654 {
1655 	struct ether_header *eh;
1656 	struct ether_vlan_header *evh;
1657 	uint16_t eth_type;
1658 	int offset, error;
1659 
1660 	eh = mtod(m, struct ether_header *);
1661 	eth_type = ntohs(eh->ether_type);
1662 	if (eth_type == ETHERTYPE_VLAN) {
1663 		/* BMV: We should handle nested VLAN tags too. */
1664 		evh = mtod(m, struct ether_vlan_header *);
1665 		eth_type = ntohs(evh->evl_proto);
1666 		offset = sizeof(struct ether_vlan_header);
1667 	} else
1668 		offset = sizeof(struct ether_header);
1669 
1670 	if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
1671 		error = ptnet_rx_csum_by_offset(m, eth_type, offset, hdr);
1672 	else
1673 		error = ptnet_rx_csum_by_parse(m, eth_type, offset, hdr);
1674 
1675 	return (error);
1676 }
1677 /* End of offloading-related functions to be shared with vtnet. */
1678 
1679 static inline void
1680 ptnet_sync_tail(struct ptnet_csb_hg *pthg, struct netmap_kring *kring)
1681 {
1682 	struct netmap_ring *ring = kring->ring;
1683 
1684 	/* Update hwcur and hwtail as known by the host. */
1685         ptnetmap_guest_read_kring_csb(pthg, kring);
1686 
1687 	/* nm_sync_finalize */
1688 	ring->tail = kring->rtail = kring->nr_hwtail;
1689 }
1690 
1691 static void
1692 ptnet_ring_update(struct ptnet_queue *pq, struct netmap_kring *kring,
1693 		  unsigned int head, unsigned int sync_flags)
1694 {
1695 	struct netmap_ring *ring = kring->ring;
1696 	struct ptnet_csb_gh *ptgh = pq->ptgh;
1697 	struct ptnet_csb_hg *pthg = pq->pthg;
1698 
1699 	/* Some packets have been pushed to the netmap ring. We have
1700 	 * to tell the host to process the new packets, updating cur
1701 	 * and head in the CSB. */
1702 	ring->head = ring->cur = head;
1703 
1704 	/* Mimic nm_txsync_prologue/nm_rxsync_prologue. */
1705 	kring->rcur = kring->rhead = head;
1706 
1707 	ptnetmap_guest_write_kring_csb(ptgh, kring->rcur, kring->rhead);
1708 
1709 	/* Kick the host if needed. */
1710 	if (NM_ACCESS_ONCE(pthg->host_need_kick)) {
1711 		ptgh->sync_flags = sync_flags;
1712 		ptnet_kick(pq);
1713 	}
1714 }
1715 
1716 #define PTNET_TX_NOSPACE(_h, _k, _min)	\
1717 	((((_h) < (_k)->rtail) ? 0 : (_k)->nkr_num_slots) + \
1718 		(_k)->rtail - (_h)) < (_min)
1719 
1720 /* This function may be called by the network stack, or by
1721  * by the taskqueue thread. */
1722 static int
1723 ptnet_drain_transmit_queue(struct ptnet_queue *pq, unsigned int budget,
1724 			   bool may_resched)
1725 {
1726 	struct ptnet_softc *sc = pq->sc;
1727 	bool have_vnet_hdr = sc->vnet_hdr_len;
1728 	struct netmap_adapter *na = &sc->ptna->dr.up;
1729 	if_t ifp = sc->ifp;
1730 	unsigned int batch_count = 0;
1731 	struct ptnet_csb_gh *ptgh;
1732 	struct ptnet_csb_hg *pthg;
1733 	struct netmap_kring *kring;
1734 	struct netmap_ring *ring;
1735 	struct netmap_slot *slot;
1736 	unsigned int count = 0;
1737 	unsigned int minspace;
1738 	unsigned int head;
1739 	unsigned int lim;
1740 	struct mbuf *mhead;
1741 	struct mbuf *mf;
1742 	int nmbuf_bytes;
1743 	uint8_t *nmbuf;
1744 
1745 	if (!PTNET_Q_TRYLOCK(pq)) {
1746 		/* We failed to acquire the lock, schedule the taskqueue. */
1747 		RD(1, "Deferring TX work");
1748 		if (may_resched) {
1749 			taskqueue_enqueue(pq->taskq, &pq->task);
1750 		}
1751 
1752 		return 0;
1753 	}
1754 
1755 	if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1756 		PTNET_Q_UNLOCK(pq);
1757 		RD(1, "Interface is down");
1758 		return ENETDOWN;
1759 	}
1760 
1761 	ptgh = pq->ptgh;
1762 	pthg = pq->pthg;
1763 	kring = na->tx_rings[pq->kring_id];
1764 	ring = kring->ring;
1765 	lim = kring->nkr_num_slots - 1;
1766 	head = ring->head;
1767 	minspace = sc->min_tx_space;
1768 
1769 	while (count < budget) {
1770 		if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1771 			/* We ran out of slot, let's see if the host has
1772 			 * freed up some, by reading hwcur and hwtail from
1773 			 * the CSB. */
1774 			ptnet_sync_tail(pthg, kring);
1775 
1776 			if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1777 				/* Still no slots available. Reactivate the
1778 				 * interrupts so that we can be notified
1779 				 * when some free slots are made available by
1780 				 * the host. */
1781 				ptgh->guest_need_kick = 1;
1782 
1783 				/* Double-check. */
1784 				ptnet_sync_tail(pthg, kring);
1785 				if (likely(PTNET_TX_NOSPACE(head, kring,
1786 							    minspace))) {
1787 					break;
1788 				}
1789 
1790 				RD(1, "Found more slots by doublecheck");
1791 				/* More slots were freed before reactivating
1792 				 * the interrupts. */
1793 				ptgh->guest_need_kick = 0;
1794 			}
1795 		}
1796 
1797 		mhead = drbr_peek(ifp, pq->bufring);
1798 		if (!mhead) {
1799 			break;
1800 		}
1801 
1802 		/* Initialize transmission state variables. */
1803 		slot = ring->slot + head;
1804 		nmbuf = NMB(na, slot);
1805 		nmbuf_bytes = 0;
1806 
1807 		/* If needed, prepare the virtio-net header at the beginning
1808 		 * of the first slot. */
1809 		if (have_vnet_hdr) {
1810 			struct virtio_net_hdr *vh =
1811 					(struct virtio_net_hdr *)nmbuf;
1812 
1813 			/* For performance, we could replace this memset() with
1814 			 * two 8-bytes-wide writes. */
1815 			memset(nmbuf, 0, PTNET_HDR_SIZE);
1816 			if (mhead->m_pkthdr.csum_flags & PTNET_ALL_OFFLOAD) {
1817 				mhead = ptnet_tx_offload(ifp, mhead, false,
1818 							 vh);
1819 				if (unlikely(!mhead)) {
1820 					/* Packet dropped because errors
1821 					 * occurred while preparing the vnet
1822 					 * header. Let's go ahead with the next
1823 					 * packet. */
1824 					pq->stats.errors ++;
1825 					drbr_advance(ifp, pq->bufring);
1826 					continue;
1827 				}
1828 			}
1829 			ND(1, "%s: [csum_flags %lX] vnet hdr: flags %x "
1830 			      "csum_start %u csum_ofs %u hdr_len = %u "
1831 			      "gso_size %u gso_type %x", __func__,
1832 			      mhead->m_pkthdr.csum_flags, vh->flags,
1833 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
1834 			      vh->gso_size, vh->gso_type);
1835 
1836 			nmbuf += PTNET_HDR_SIZE;
1837 			nmbuf_bytes += PTNET_HDR_SIZE;
1838 		}
1839 
1840 		for (mf = mhead; mf; mf = mf->m_next) {
1841 			uint8_t *mdata = mf->m_data;
1842 			int mlen = mf->m_len;
1843 
1844 			for (;;) {
1845 				int copy = NETMAP_BUF_SIZE(na) - nmbuf_bytes;
1846 
1847 				if (mlen < copy) {
1848 					copy = mlen;
1849 				}
1850 				memcpy(nmbuf, mdata, copy);
1851 
1852 				mdata += copy;
1853 				mlen -= copy;
1854 				nmbuf += copy;
1855 				nmbuf_bytes += copy;
1856 
1857 				if (!mlen) {
1858 					break;
1859 				}
1860 
1861 				slot->len = nmbuf_bytes;
1862 				slot->flags = NS_MOREFRAG;
1863 
1864 				head = nm_next(head, lim);
1865 				KASSERT(head != ring->tail,
1866 					("Unexpectedly run out of TX space"));
1867 				slot = ring->slot + head;
1868 				nmbuf = NMB(na, slot);
1869 				nmbuf_bytes = 0;
1870 			}
1871 		}
1872 
1873 		/* Complete last slot and update head. */
1874 		slot->len = nmbuf_bytes;
1875 		slot->flags = 0;
1876 		head = nm_next(head, lim);
1877 
1878 		/* Consume the packet just processed. */
1879 		drbr_advance(ifp, pq->bufring);
1880 
1881 		/* Copy the packet to listeners. */
1882 		ETHER_BPF_MTAP(ifp, mhead);
1883 
1884 		pq->stats.packets ++;
1885 		pq->stats.bytes += mhead->m_pkthdr.len;
1886 		if (mhead->m_flags & M_MCAST) {
1887 			pq->stats.mcasts ++;
1888 		}
1889 
1890 		m_freem(mhead);
1891 
1892 		count ++;
1893 		if (++batch_count == PTNET_TX_BATCH) {
1894 			ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1895 			batch_count = 0;
1896 		}
1897 	}
1898 
1899 	if (batch_count) {
1900 		ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1901 	}
1902 
1903 	if (count >= budget && may_resched) {
1904 		DBG(RD(1, "out of budget: resched, %d mbufs pending\n",
1905 					drbr_inuse(ifp, pq->bufring)));
1906 		taskqueue_enqueue(pq->taskq, &pq->task);
1907 	}
1908 
1909 	PTNET_Q_UNLOCK(pq);
1910 
1911 	return count;
1912 }
1913 
1914 static int
1915 ptnet_transmit(if_t ifp, struct mbuf *m)
1916 {
1917 	struct ptnet_softc *sc = if_getsoftc(ifp);
1918 	struct ptnet_queue *pq;
1919 	unsigned int queue_idx;
1920 	int err;
1921 
1922 	DBG(device_printf(sc->dev, "transmit %p\n", m));
1923 
1924 	/* Insert 802.1Q header if needed. */
1925 	if (m->m_flags & M_VLANTAG) {
1926 		m = ether_vlanencap(m, m->m_pkthdr.ether_vtag);
1927 		if (m == NULL) {
1928 			return ENOBUFS;
1929 		}
1930 		m->m_flags &= ~M_VLANTAG;
1931 	}
1932 
1933 	/* Get the flow-id if available. */
1934 	queue_idx = (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) ?
1935 		    m->m_pkthdr.flowid : curcpu;
1936 
1937 	if (unlikely(queue_idx >= sc->num_tx_rings)) {
1938 		queue_idx %= sc->num_tx_rings;
1939 	}
1940 
1941 	pq = sc->queues + queue_idx;
1942 
1943 	err = drbr_enqueue(ifp, pq->bufring, m);
1944 	if (err) {
1945 		/* ENOBUFS when the bufring is full */
1946 		RD(1, "%s: drbr_enqueue() failed %d\n",
1947 			__func__, err);
1948 		pq->stats.errors ++;
1949 		return err;
1950 	}
1951 
1952 	if (ifp->if_capenable & IFCAP_POLLING) {
1953 		/* If polling is on, the transmit queues will be
1954 		 * drained by the poller. */
1955 		return 0;
1956 	}
1957 
1958 	err = ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
1959 
1960 	return (err < 0) ? err : 0;
1961 }
1962 
1963 static unsigned int
1964 ptnet_rx_discard(struct netmap_kring *kring, unsigned int head)
1965 {
1966 	struct netmap_ring *ring = kring->ring;
1967 	struct netmap_slot *slot = ring->slot + head;
1968 
1969 	for (;;) {
1970 		head = nm_next(head, kring->nkr_num_slots - 1);
1971 		if (!(slot->flags & NS_MOREFRAG) || head == ring->tail) {
1972 			break;
1973 		}
1974 		slot = ring->slot + head;
1975 	}
1976 
1977 	return head;
1978 }
1979 
1980 static inline struct mbuf *
1981 ptnet_rx_slot(struct mbuf *mtail, uint8_t *nmbuf, unsigned int nmbuf_len)
1982 {
1983 	uint8_t *mdata = mtod(mtail, uint8_t *) + mtail->m_len;
1984 
1985 	do {
1986 		unsigned int copy;
1987 
1988 		if (mtail->m_len == MCLBYTES) {
1989 			struct mbuf *mf;
1990 
1991 			mf = m_getcl(M_NOWAIT, MT_DATA, 0);
1992 			if (unlikely(!mf)) {
1993 				return NULL;
1994 			}
1995 
1996 			mtail->m_next = mf;
1997 			mtail = mf;
1998 			mdata = mtod(mtail, uint8_t *);
1999 			mtail->m_len = 0;
2000 		}
2001 
2002 		copy = MCLBYTES - mtail->m_len;
2003 		if (nmbuf_len < copy) {
2004 			copy = nmbuf_len;
2005 		}
2006 
2007 		memcpy(mdata, nmbuf, copy);
2008 
2009 		nmbuf += copy;
2010 		nmbuf_len -= copy;
2011 		mdata += copy;
2012 		mtail->m_len += copy;
2013 	} while (nmbuf_len);
2014 
2015 	return mtail;
2016 }
2017 
2018 static int
2019 ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget, bool may_resched)
2020 {
2021 	struct ptnet_softc *sc = pq->sc;
2022 	bool have_vnet_hdr = sc->vnet_hdr_len;
2023 	struct ptnet_csb_gh *ptgh = pq->ptgh;
2024 	struct ptnet_csb_hg *pthg = pq->pthg;
2025 	struct netmap_adapter *na = &sc->ptna->dr.up;
2026 	struct netmap_kring *kring = na->rx_rings[pq->kring_id];
2027 	struct netmap_ring *ring = kring->ring;
2028 	unsigned int const lim = kring->nkr_num_slots - 1;
2029 	unsigned int batch_count = 0;
2030 	if_t ifp = sc->ifp;
2031 	unsigned int count = 0;
2032 	uint32_t head;
2033 
2034 	PTNET_Q_LOCK(pq);
2035 
2036 	if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
2037 		goto unlock;
2038 	}
2039 
2040 	kring->nr_kflags &= ~NKR_PENDINTR;
2041 
2042 	head = ring->head;
2043 	while (count < budget) {
2044 		uint32_t prev_head = head;
2045 		struct mbuf *mhead, *mtail;
2046 		struct virtio_net_hdr *vh;
2047 		struct netmap_slot *slot;
2048 		unsigned int nmbuf_len;
2049 		uint8_t *nmbuf;
2050 		int deliver = 1; /* the mbuf to the network stack. */
2051 host_sync:
2052 		if (head == ring->tail) {
2053 			/* We ran out of slot, let's see if the host has
2054 			 * added some, by reading hwcur and hwtail from
2055 			 * the CSB. */
2056 			ptnet_sync_tail(pthg, kring);
2057 
2058 			if (head == ring->tail) {
2059 				/* Still no slots available. Reactivate
2060 				 * interrupts as they were disabled by the
2061 				 * host thread right before issuing the
2062 				 * last interrupt. */
2063 				ptgh->guest_need_kick = 1;
2064 
2065 				/* Double-check. */
2066 				ptnet_sync_tail(pthg, kring);
2067 				if (likely(head == ring->tail)) {
2068 					break;
2069 				}
2070 				ptgh->guest_need_kick = 0;
2071 			}
2072 		}
2073 
2074 		/* Initialize ring state variables, possibly grabbing the
2075 		 * virtio-net header. */
2076 		slot = ring->slot + head;
2077 		nmbuf = NMB(na, slot);
2078 		nmbuf_len = slot->len;
2079 
2080 		vh = (struct virtio_net_hdr *)nmbuf;
2081 		if (have_vnet_hdr) {
2082 			if (unlikely(nmbuf_len < PTNET_HDR_SIZE)) {
2083 				/* There is no good reason why host should
2084 				 * put the header in multiple netmap slots.
2085 				 * If this is the case, discard. */
2086 				RD(1, "Fragmented vnet-hdr: dropping");
2087 				head = ptnet_rx_discard(kring, head);
2088 				pq->stats.iqdrops ++;
2089 				deliver = 0;
2090 				goto skip;
2091 			}
2092 			ND(1, "%s: vnet hdr: flags %x csum_start %u "
2093 			      "csum_ofs %u hdr_len = %u gso_size %u "
2094 			      "gso_type %x", __func__, vh->flags,
2095 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
2096 			      vh->gso_size, vh->gso_type);
2097 			nmbuf += PTNET_HDR_SIZE;
2098 			nmbuf_len -= PTNET_HDR_SIZE;
2099 		}
2100 
2101 		/* Allocate the head of a new mbuf chain.
2102 		 * We use m_getcl() to allocate an mbuf with standard cluster
2103 		 * size (MCLBYTES). In the future we could use m_getjcl()
2104 		 * to choose different sizes. */
2105 		mhead = mtail = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
2106 		if (unlikely(mhead == NULL)) {
2107 			device_printf(sc->dev, "%s: failed to allocate mbuf "
2108 				      "head\n", __func__);
2109 			pq->stats.errors ++;
2110 			break;
2111 		}
2112 
2113 		/* Initialize the mbuf state variables. */
2114 		mhead->m_pkthdr.len = nmbuf_len;
2115 		mtail->m_len = 0;
2116 
2117 		/* Scan all the netmap slots containing the current packet. */
2118 		for (;;) {
2119 			DBG(device_printf(sc->dev, "%s: h %u t %u rcv frag "
2120 					  "len %u, flags %u\n", __func__,
2121 					  head, ring->tail, slot->len,
2122 					  slot->flags));
2123 
2124 			mtail = ptnet_rx_slot(mtail, nmbuf, nmbuf_len);
2125 			if (unlikely(!mtail)) {
2126 				/* Ouch. We ran out of memory while processing
2127 				 * a packet. We have to restore the previous
2128 				 * head position, free the mbuf chain, and
2129 				 * schedule the taskqueue to give the packet
2130 				 * another chance. */
2131 				device_printf(sc->dev, "%s: failed to allocate"
2132 					" mbuf frag, reset head %u --> %u\n",
2133 					__func__, head, prev_head);
2134 				head = prev_head;
2135 				m_freem(mhead);
2136 				pq->stats.errors ++;
2137 				if (may_resched) {
2138 					taskqueue_enqueue(pq->taskq,
2139 							  &pq->task);
2140 				}
2141 				goto escape;
2142 			}
2143 
2144 			/* We have to increment head irrespective of the
2145 			 * NS_MOREFRAG being set or not. */
2146 			head = nm_next(head, lim);
2147 
2148 			if (!(slot->flags & NS_MOREFRAG)) {
2149 				break;
2150 			}
2151 
2152 			if (unlikely(head == ring->tail)) {
2153 				/* The very last slot prepared by the host has
2154 				 * the NS_MOREFRAG set. Drop it and continue
2155 				 * the outer cycle (to do the double-check). */
2156 				RD(1, "Incomplete packet: dropping");
2157 				m_freem(mhead);
2158 				pq->stats.iqdrops ++;
2159 				goto host_sync;
2160 			}
2161 
2162 			slot = ring->slot + head;
2163 			nmbuf = NMB(na, slot);
2164 			nmbuf_len = slot->len;
2165 			mhead->m_pkthdr.len += nmbuf_len;
2166 		}
2167 
2168 		mhead->m_pkthdr.rcvif = ifp;
2169 		mhead->m_pkthdr.csum_flags = 0;
2170 
2171 		/* Store the queue idx in the packet header. */
2172 		mhead->m_pkthdr.flowid = pq->kring_id;
2173 		M_HASHTYPE_SET(mhead, M_HASHTYPE_OPAQUE);
2174 
2175 		if (ifp->if_capenable & IFCAP_VLAN_HWTAGGING) {
2176 			struct ether_header *eh;
2177 
2178 			eh = mtod(mhead, struct ether_header *);
2179 			if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
2180 				ptnet_vlan_tag_remove(mhead);
2181 				/*
2182 				 * With the 802.1Q header removed, update the
2183 				 * checksum starting location accordingly.
2184 				 */
2185 				if (vh->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
2186 					vh->csum_start -= ETHER_VLAN_ENCAP_LEN;
2187 			}
2188 		}
2189 
2190 		if (have_vnet_hdr && (vh->flags & (VIRTIO_NET_HDR_F_NEEDS_CSUM
2191 					| VIRTIO_NET_HDR_F_DATA_VALID))) {
2192 			if (unlikely(ptnet_rx_csum(mhead, vh))) {
2193 				m_freem(mhead);
2194 				RD(1, "Csum offload error: dropping");
2195 				pq->stats.iqdrops ++;
2196 				deliver = 0;
2197 			}
2198 		}
2199 
2200 skip:
2201 		count ++;
2202 		if (++batch_count >= PTNET_RX_BATCH) {
2203 			/* Some packets have been (or will be) pushed to the network
2204 			 * stack. We need to update the CSB to tell the host about
2205 			 * the new ring->cur and ring->head (RX buffer refill). */
2206 			ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
2207 			batch_count = 0;
2208 		}
2209 
2210 		if (likely(deliver))  {
2211 			pq->stats.packets ++;
2212 			pq->stats.bytes += mhead->m_pkthdr.len;
2213 
2214 			PTNET_Q_UNLOCK(pq);
2215 			(*ifp->if_input)(ifp, mhead);
2216 			PTNET_Q_LOCK(pq);
2217 			/* The ring->head index (and related indices) are
2218 			 * updated under pq lock by ptnet_ring_update().
2219 			 * Since we dropped the lock to call if_input(), we
2220 			 * must reload ring->head and restart processing the
2221 			 * ring from there. */
2222 			head = ring->head;
2223 
2224 			if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
2225 				/* The interface has gone down while we didn't
2226 				 * have the lock. Stop any processing and exit. */
2227 				goto unlock;
2228 			}
2229 		}
2230 	}
2231 escape:
2232 	if (batch_count) {
2233 		ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
2234 
2235 	}
2236 
2237 	if (count >= budget && may_resched) {
2238 		/* If we ran out of budget or the double-check found new
2239 		 * slots to process, schedule the taskqueue. */
2240 		DBG(RD(1, "out of budget: resched h %u t %u\n",
2241 					head, ring->tail));
2242 		taskqueue_enqueue(pq->taskq, &pq->task);
2243 	}
2244 unlock:
2245 	PTNET_Q_UNLOCK(pq);
2246 
2247 	return count;
2248 }
2249 
2250 static void
2251 ptnet_rx_task(void *context, int pending)
2252 {
2253 	struct ptnet_queue *pq = context;
2254 
2255 	DBG(RD(1, "%s: pq #%u\n", __func__, pq->kring_id));
2256 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
2257 }
2258 
2259 static void
2260 ptnet_tx_task(void *context, int pending)
2261 {
2262 	struct ptnet_queue *pq = context;
2263 
2264 	DBG(RD(1, "%s: pq #%u\n", __func__, pq->kring_id));
2265 	ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
2266 }
2267 
2268 #ifdef DEVICE_POLLING
2269 /* We don't need to handle differently POLL_AND_CHECK_STATUS and
2270  * POLL_ONLY, since we don't have an Interrupt Status Register. */
2271 static int
2272 ptnet_poll(if_t ifp, enum poll_cmd cmd, int budget)
2273 {
2274 	struct ptnet_softc *sc = if_getsoftc(ifp);
2275 	unsigned int queue_budget;
2276 	unsigned int count = 0;
2277 	bool borrow = false;
2278 	int i;
2279 
2280 	KASSERT(sc->num_rings > 0, ("Found no queues in while polling ptnet"));
2281 	queue_budget = MAX(budget / sc->num_rings, 1);
2282 	RD(1, "Per-queue budget is %d", queue_budget);
2283 
2284 	while (budget) {
2285 		unsigned int rcnt = 0;
2286 
2287 		for (i = 0; i < sc->num_rings; i++) {
2288 			struct ptnet_queue *pq = sc->queues + i;
2289 
2290 			if (borrow) {
2291 				queue_budget = MIN(queue_budget, budget);
2292 				if (queue_budget == 0) {
2293 					break;
2294 				}
2295 			}
2296 
2297 			if (i < sc->num_tx_rings) {
2298 				rcnt += ptnet_drain_transmit_queue(pq,
2299 						   queue_budget, false);
2300 			} else {
2301 				rcnt += ptnet_rx_eof(pq, queue_budget,
2302 						      false);
2303 			}
2304 		}
2305 
2306 		if (!rcnt) {
2307 			/* A scan of the queues gave no result, we can
2308 			 * stop here. */
2309 			break;
2310 		}
2311 
2312 		if (rcnt > budget) {
2313 			/* This may happen when initial budget < sc->num_rings,
2314 			 * since one packet budget is given to each queue
2315 			 * anyway. Just pretend we didn't eat "so much". */
2316 			rcnt = budget;
2317 		}
2318 		count += rcnt;
2319 		budget -= rcnt;
2320 		borrow = true;
2321 	}
2322 
2323 
2324 	return count;
2325 }
2326 #endif /* DEVICE_POLLING */
2327