xref: /freebsd/sys/dev/netmap/if_ptnet.c (revision a7623790fb345e6dc986dfd31df0ace115e6f2e4)
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 
72 #include <machine/bus.h>
73 #include <machine/resource.h>
74 #include <sys/bus.h>
75 #include <sys/rman.h>
76 
77 #include <dev/pci/pcivar.h>
78 #include <dev/pci/pcireg.h>
79 
80 #include "opt_inet.h"
81 #include "opt_inet6.h"
82 
83 #include <sys/selinfo.h>
84 #include <net/netmap.h>
85 #include <dev/netmap/netmap_kern.h>
86 #include <net/netmap_virt.h>
87 #include <dev/netmap/netmap_mem2.h>
88 #include <dev/virtio/network/virtio_net.h>
89 
90 #ifdef WITH_PTNETMAP
91 
92 #ifndef INET
93 #error "INET not defined, cannot support offloadings"
94 #endif
95 
96 #if __FreeBSD_version >= 1100000
97 static uint64_t	ptnet_get_counter(if_t, ift_counter);
98 #else
99 typedef struct ifnet *if_t;
100 #define if_getsoftc(_ifp)   (_ifp)->if_softc
101 #endif
102 
103 //#define PTNETMAP_STATS
104 //#define DEBUG
105 #ifdef DEBUG
106 #define DBG(x) x
107 #else   /* !DEBUG */
108 #define DBG(x)
109 #endif  /* !DEBUG */
110 
111 extern int ptnet_vnet_hdr; /* Tunable parameter */
112 
113 struct ptnet_softc;
114 
115 struct ptnet_queue_stats {
116 	uint64_t	packets; /* if_[io]packets */
117 	uint64_t	bytes;	 /* if_[io]bytes */
118 	uint64_t	errors;	 /* if_[io]errors */
119 	uint64_t	iqdrops; /* if_iqdrops */
120 	uint64_t	mcasts;  /* if_[io]mcasts */
121 #ifdef PTNETMAP_STATS
122 	uint64_t	intrs;
123 	uint64_t	kicks;
124 #endif /* PTNETMAP_STATS */
125 };
126 
127 struct ptnet_queue {
128 	struct ptnet_softc		*sc;
129 	struct				resource *irq;
130 	void				*cookie;
131 	int				kring_id;
132 	struct nm_csb_atok		*atok;
133 	struct nm_csb_ktoa		*ktoa;
134 	unsigned int			kick;
135 	struct mtx			lock;
136 	struct buf_ring			*bufring; /* for TX queues */
137 	struct ptnet_queue_stats	stats;
138 #ifdef PTNETMAP_STATS
139 	struct ptnet_queue_stats	last_stats;
140 #endif /* PTNETMAP_STATS */
141 	struct taskqueue		*taskq;
142 	struct task			task;
143 	char				lock_name[16];
144 };
145 
146 #define PTNET_Q_LOCK(_pq)	mtx_lock(&(_pq)->lock)
147 #define PTNET_Q_TRYLOCK(_pq)	mtx_trylock(&(_pq)->lock)
148 #define PTNET_Q_UNLOCK(_pq)	mtx_unlock(&(_pq)->lock)
149 
150 struct ptnet_softc {
151 	device_t		dev;
152 	if_t			ifp;
153 	struct ifmedia		media;
154 	struct mtx		lock;
155 	char			lock_name[16];
156 	char			hwaddr[ETHER_ADDR_LEN];
157 
158 	/* Mirror of PTFEAT register. */
159 	uint32_t		ptfeatures;
160 	unsigned int		vnet_hdr_len;
161 
162 	/* PCI BARs support. */
163 	struct resource		*iomem;
164 	struct resource		*msix_mem;
165 
166 	unsigned int		num_rings;
167 	unsigned int		num_tx_rings;
168 	struct ptnet_queue	*queues;
169 	struct ptnet_queue	*rxqueues;
170 	struct nm_csb_atok	*csb_gh;
171 	struct nm_csb_ktoa	*csb_hg;
172 
173 	unsigned int		min_tx_space;
174 
175 	struct netmap_pt_guest_adapter *ptna;
176 
177 	struct callout		tick;
178 #ifdef PTNETMAP_STATS
179 	struct timeval		last_ts;
180 #endif /* PTNETMAP_STATS */
181 };
182 
183 #define PTNET_CORE_LOCK(_sc)	mtx_lock(&(_sc)->lock)
184 #define PTNET_CORE_UNLOCK(_sc)	mtx_unlock(&(_sc)->lock)
185 
186 static int	ptnet_probe(device_t);
187 static int	ptnet_attach(device_t);
188 static int	ptnet_detach(device_t);
189 static int	ptnet_suspend(device_t);
190 static int	ptnet_resume(device_t);
191 static int	ptnet_shutdown(device_t);
192 
193 static void	ptnet_init(void *opaque);
194 static int	ptnet_ioctl(if_t ifp, u_long cmd, caddr_t data);
195 static int	ptnet_init_locked(struct ptnet_softc *sc);
196 static int	ptnet_stop(struct ptnet_softc *sc);
197 static int	ptnet_transmit(if_t ifp, struct mbuf *m);
198 static int	ptnet_drain_transmit_queue(struct ptnet_queue *pq,
199 					   unsigned int budget,
200 					   bool may_resched);
201 static void	ptnet_qflush(if_t ifp);
202 static void	ptnet_tx_task(void *context, int pending);
203 
204 static int	ptnet_media_change(if_t ifp);
205 static void	ptnet_media_status(if_t ifp, struct ifmediareq *ifmr);
206 #ifdef PTNETMAP_STATS
207 static void	ptnet_tick(void *opaque);
208 #endif
209 
210 static int	ptnet_irqs_init(struct ptnet_softc *sc);
211 static void	ptnet_irqs_fini(struct ptnet_softc *sc);
212 
213 static uint32_t ptnet_nm_ptctl(struct ptnet_softc *sc, uint32_t cmd);
214 static int      ptnet_nm_config(struct netmap_adapter *na,
215 				struct nm_config_info *info);
216 static void	ptnet_update_vnet_hdr(struct ptnet_softc *sc);
217 static int	ptnet_nm_register(struct netmap_adapter *na, int onoff);
218 static int	ptnet_nm_txsync(struct netmap_kring *kring, int flags);
219 static int	ptnet_nm_rxsync(struct netmap_kring *kring, int flags);
220 static void	ptnet_nm_intr(struct netmap_adapter *na, int onoff);
221 
222 static void	ptnet_tx_intr(void *opaque);
223 static void	ptnet_rx_intr(void *opaque);
224 
225 static unsigned	ptnet_rx_discard(struct netmap_kring *kring,
226 				 unsigned int head);
227 static int	ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget,
228 			     bool may_resched);
229 static void	ptnet_rx_task(void *context, int pending);
230 
231 #ifdef DEVICE_POLLING
232 static poll_handler_t ptnet_poll;
233 #endif
234 
235 static device_method_t ptnet_methods[] = {
236 	DEVMETHOD(device_probe,			ptnet_probe),
237 	DEVMETHOD(device_attach,		ptnet_attach),
238 	DEVMETHOD(device_detach,		ptnet_detach),
239 	DEVMETHOD(device_suspend,		ptnet_suspend),
240 	DEVMETHOD(device_resume,		ptnet_resume),
241 	DEVMETHOD(device_shutdown,		ptnet_shutdown),
242 	DEVMETHOD_END
243 };
244 
245 static driver_t ptnet_driver = {
246 	"ptnet",
247 	ptnet_methods,
248 	sizeof(struct ptnet_softc)
249 };
250 
251 /* We use (SI_ORDER_MIDDLE+2) here, see DEV_MODULE_ORDERED() invocation. */
252 static devclass_t ptnet_devclass;
253 DRIVER_MODULE_ORDERED(ptnet, pci, ptnet_driver, ptnet_devclass,
254 		      NULL, NULL, SI_ORDER_MIDDLE + 2);
255 
256 static int
257 ptnet_probe(device_t dev)
258 {
259 	if (pci_get_vendor(dev) != PTNETMAP_PCI_VENDOR_ID ||
260 		pci_get_device(dev) != PTNETMAP_PCI_NETIF_ID) {
261 		return (ENXIO);
262 	}
263 
264 	device_set_desc(dev, "ptnet network adapter");
265 
266 	return (BUS_PROBE_DEFAULT);
267 }
268 
269 static inline void ptnet_kick(struct ptnet_queue *pq)
270 {
271 #ifdef PTNETMAP_STATS
272 	pq->stats.kicks ++;
273 #endif /* PTNETMAP_STATS */
274 	bus_write_4(pq->sc->iomem, pq->kick, 0);
275 }
276 
277 #define PTNET_BUF_RING_SIZE	4096
278 #define PTNET_RX_BUDGET		512
279 #define PTNET_RX_BATCH		1
280 #define PTNET_TX_BUDGET		512
281 #define PTNET_TX_BATCH		64
282 #define PTNET_HDR_SIZE		sizeof(struct virtio_net_hdr_mrg_rxbuf)
283 #define PTNET_MAX_PKT_SIZE	65536
284 
285 #define PTNET_CSUM_OFFLOAD	(CSUM_TCP | CSUM_UDP)
286 #define PTNET_CSUM_OFFLOAD_IPV6	(CSUM_TCP_IPV6 | CSUM_UDP_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 nm_csb_atok) > 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 nm_csb_ktoa *)(((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->atok = sc->csb_gh + i;
383 		pq->ktoa = 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 /* Stop host sync-kloop if it was running. */
509 static void
510 ptnet_device_shutdown(struct ptnet_softc *sc)
511 {
512 	ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_DELETE);
513 	bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAH, 0);
514 	bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAL, 0);
515 	bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAH, 0);
516 	bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAL, 0);
517 }
518 
519 static int
520 ptnet_detach(device_t dev)
521 {
522 	struct ptnet_softc *sc = device_get_softc(dev);
523 	int i;
524 
525 	ptnet_device_shutdown(sc);
526 
527 #ifdef DEVICE_POLLING
528 	if (sc->ifp->if_capenable & IFCAP_POLLING) {
529 		ether_poll_deregister(sc->ifp);
530 	}
531 #endif
532 	callout_drain(&sc->tick);
533 
534 	if (sc->queues) {
535 		/* Drain taskqueues before calling if_detach. */
536 		for (i = 0; i < sc->num_rings; i++) {
537 			struct ptnet_queue *pq = sc->queues + i;
538 
539 			if (pq->taskq) {
540 				taskqueue_drain(pq->taskq, &pq->task);
541 			}
542 		}
543 	}
544 
545 	if (sc->ifp) {
546 		ether_ifdetach(sc->ifp);
547 
548 		/* Uninitialize netmap adapters for this device. */
549 		netmap_detach(sc->ifp);
550 
551 		ifmedia_removeall(&sc->media);
552 		if_free(sc->ifp);
553 		sc->ifp = NULL;
554 	}
555 
556 	ptnet_irqs_fini(sc);
557 
558 	if (sc->csb_gh) {
559 		contigfree(sc->csb_gh, 2*PAGE_SIZE, M_DEVBUF);
560 		sc->csb_gh = NULL;
561 		sc->csb_hg = NULL;
562 	}
563 
564 	if (sc->queues) {
565 		for (i = 0; i < sc->num_rings; i++) {
566 			struct ptnet_queue *pq = sc->queues + i;
567 
568 			if (mtx_initialized(&pq->lock)) {
569 				mtx_destroy(&pq->lock);
570 			}
571 			if (pq->bufring != NULL) {
572 				buf_ring_free(pq->bufring, M_DEVBUF);
573 			}
574 		}
575 		free(sc->queues, M_DEVBUF);
576 		sc->queues = NULL;
577 	}
578 
579 	if (sc->iomem) {
580 		bus_release_resource(dev, SYS_RES_IOPORT,
581 				     PCIR_BAR(PTNETMAP_IO_PCI_BAR), sc->iomem);
582 		sc->iomem = NULL;
583 	}
584 
585 	mtx_destroy(&sc->lock);
586 
587 	device_printf(dev, "%s() completed\n", __func__);
588 
589 	return (0);
590 }
591 
592 static int
593 ptnet_suspend(device_t dev)
594 {
595 	struct ptnet_softc *sc = device_get_softc(dev);
596 
597 	(void)sc;
598 
599 	return (0);
600 }
601 
602 static int
603 ptnet_resume(device_t dev)
604 {
605 	struct ptnet_softc *sc = device_get_softc(dev);
606 
607 	(void)sc;
608 
609 	return (0);
610 }
611 
612 static int
613 ptnet_shutdown(device_t dev)
614 {
615 	struct ptnet_softc *sc = device_get_softc(dev);
616 
617 	ptnet_device_shutdown(sc);
618 
619 	return (0);
620 }
621 
622 static int
623 ptnet_irqs_init(struct ptnet_softc *sc)
624 {
625 	int rid = PCIR_BAR(PTNETMAP_MSIX_PCI_BAR);
626 	int nvecs = sc->num_rings;
627 	device_t dev = sc->dev;
628 	int err = ENOSPC;
629 	int cpu_cur;
630 	int i;
631 
632 	if (pci_find_cap(dev, PCIY_MSIX, NULL) != 0)  {
633 		device_printf(dev, "Could not find MSI-X capability\n");
634 		return (ENXIO);
635 	}
636 
637 	sc->msix_mem = bus_alloc_resource_any(dev, SYS_RES_MEMORY,
638 					      &rid, RF_ACTIVE);
639 	if (sc->msix_mem == NULL) {
640 		device_printf(dev, "Failed to allocate MSIX PCI BAR\n");
641 		return (ENXIO);
642 	}
643 
644 	if (pci_msix_count(dev) < nvecs) {
645 		device_printf(dev, "Not enough MSI-X vectors\n");
646 		goto err_path;
647 	}
648 
649 	err = pci_alloc_msix(dev, &nvecs);
650 	if (err) {
651 		device_printf(dev, "Failed to allocate MSI-X vectors\n");
652 		goto err_path;
653 	}
654 
655 	for (i = 0; i < nvecs; i++) {
656 		struct ptnet_queue *pq = sc->queues + i;
657 
658 		rid = i + 1;
659 		pq->irq = bus_alloc_resource_any(dev, SYS_RES_IRQ, &rid,
660 						 RF_ACTIVE);
661 		if (pq->irq == NULL) {
662 			device_printf(dev, "Failed to allocate interrupt "
663 					   "for queue #%d\n", i);
664 			err = ENOSPC;
665 			goto err_path;
666 		}
667 	}
668 
669 	cpu_cur = CPU_FIRST();
670 	for (i = 0; i < nvecs; i++) {
671 		struct ptnet_queue *pq = sc->queues + i;
672 		void (*handler)(void *) = ptnet_tx_intr;
673 
674 		if (i >= sc->num_tx_rings) {
675 			handler = ptnet_rx_intr;
676 		}
677 		err = bus_setup_intr(dev, pq->irq, INTR_TYPE_NET | INTR_MPSAFE,
678 				     NULL /* intr_filter */, handler,
679 				     pq, &pq->cookie);
680 		if (err) {
681 			device_printf(dev, "Failed to register intr handler "
682 					   "for queue #%d\n", i);
683 			goto err_path;
684 		}
685 
686 		bus_describe_intr(dev, pq->irq, pq->cookie, "q%d", i);
687 #if 0
688 		bus_bind_intr(sc->dev, pq->irq, cpu_cur);
689 #endif
690 		cpu_cur = CPU_NEXT(cpu_cur);
691 	}
692 
693 	device_printf(dev, "Allocated %d MSI-X vectors\n", nvecs);
694 
695 	cpu_cur = CPU_FIRST();
696 	for (i = 0; i < nvecs; i++) {
697 		struct ptnet_queue *pq = sc->queues + i;
698 
699 		if (i < sc->num_tx_rings)
700 			TASK_INIT(&pq->task, 0, ptnet_tx_task, pq);
701 		else
702 			NET_TASK_INIT(&pq->task, 0, ptnet_rx_task, pq);
703 
704 		pq->taskq = taskqueue_create_fast("ptnet_queue", M_NOWAIT,
705 					taskqueue_thread_enqueue, &pq->taskq);
706 		taskqueue_start_threads(&pq->taskq, 1, PI_NET, "%s-pq-%d",
707 					device_get_nameunit(sc->dev), cpu_cur);
708 		cpu_cur = CPU_NEXT(cpu_cur);
709 	}
710 
711 	return 0;
712 err_path:
713 	ptnet_irqs_fini(sc);
714 	return err;
715 }
716 
717 static void
718 ptnet_irqs_fini(struct ptnet_softc *sc)
719 {
720 	device_t dev = sc->dev;
721 	int i;
722 
723 	for (i = 0; i < sc->num_rings; i++) {
724 		struct ptnet_queue *pq = sc->queues + i;
725 
726 		if (pq->taskq) {
727 			taskqueue_free(pq->taskq);
728 			pq->taskq = NULL;
729 		}
730 
731 		if (pq->cookie) {
732 			bus_teardown_intr(dev, pq->irq, pq->cookie);
733 			pq->cookie = NULL;
734 		}
735 
736 		if (pq->irq) {
737 			bus_release_resource(dev, SYS_RES_IRQ, i + 1, pq->irq);
738 			pq->irq = NULL;
739 		}
740 	}
741 
742 	if (sc->msix_mem) {
743 		pci_release_msi(dev);
744 
745 		bus_release_resource(dev, SYS_RES_MEMORY,
746 				     PCIR_BAR(PTNETMAP_MSIX_PCI_BAR),
747 				     sc->msix_mem);
748 		sc->msix_mem = NULL;
749 	}
750 }
751 
752 static void
753 ptnet_init(void *opaque)
754 {
755 	struct ptnet_softc *sc = opaque;
756 
757 	PTNET_CORE_LOCK(sc);
758 	ptnet_init_locked(sc);
759 	PTNET_CORE_UNLOCK(sc);
760 }
761 
762 static int
763 ptnet_ioctl(if_t ifp, u_long cmd, caddr_t data)
764 {
765 	struct ptnet_softc *sc = if_getsoftc(ifp);
766 	device_t dev = sc->dev;
767 	struct ifreq *ifr = (struct ifreq *)data;
768 	int mask __unused, err = 0;
769 
770 	switch (cmd) {
771 	case SIOCSIFFLAGS:
772 		device_printf(dev, "SIOCSIFFLAGS %x\n", ifp->if_flags);
773 		PTNET_CORE_LOCK(sc);
774 		if (ifp->if_flags & IFF_UP) {
775 			/* Network stack wants the iff to be up. */
776 			err = ptnet_init_locked(sc);
777 		} else {
778 			/* Network stack wants the iff to be down. */
779 			err = ptnet_stop(sc);
780 		}
781 		/* We don't need to do nothing to support IFF_PROMISC,
782 		 * since that is managed by the backend port. */
783 		PTNET_CORE_UNLOCK(sc);
784 		break;
785 
786 	case SIOCSIFCAP:
787 		device_printf(dev, "SIOCSIFCAP %x %x\n",
788 			      ifr->ifr_reqcap, ifp->if_capenable);
789 		mask = ifr->ifr_reqcap ^ ifp->if_capenable;
790 #ifdef DEVICE_POLLING
791 		if (mask & IFCAP_POLLING) {
792 			struct ptnet_queue *pq;
793 			int i;
794 
795 			if (ifr->ifr_reqcap & IFCAP_POLLING) {
796 				err = ether_poll_register(ptnet_poll, ifp);
797 				if (err) {
798 					break;
799 				}
800 				/* Stop queues and sync with taskqueues. */
801 				ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
802 				for (i = 0; i < sc->num_rings; i++) {
803 					pq = sc-> queues + i;
804 					/* Make sure the worker sees the
805 					 * IFF_DRV_RUNNING down. */
806 					PTNET_Q_LOCK(pq);
807 					pq->atok->appl_need_kick = 0;
808 					PTNET_Q_UNLOCK(pq);
809 					/* Wait for rescheduling to finish. */
810 					if (pq->taskq) {
811 						taskqueue_drain(pq->taskq,
812 								&pq->task);
813 					}
814 				}
815 				ifp->if_drv_flags |= IFF_DRV_RUNNING;
816 			} else {
817 				err = ether_poll_deregister(ifp);
818 				for (i = 0; i < sc->num_rings; i++) {
819 					pq = sc-> queues + i;
820 					PTNET_Q_LOCK(pq);
821 					pq->atok->appl_need_kick = 1;
822 					PTNET_Q_UNLOCK(pq);
823 				}
824 			}
825 		}
826 #endif  /* DEVICE_POLLING */
827 		ifp->if_capenable = ifr->ifr_reqcap;
828 		break;
829 
830 	case SIOCSIFMTU:
831 		/* We support any reasonable MTU. */
832 		if (ifr->ifr_mtu < ETHERMIN ||
833 				ifr->ifr_mtu > PTNET_MAX_PKT_SIZE) {
834 			err = EINVAL;
835 		} else {
836 			PTNET_CORE_LOCK(sc);
837 			ifp->if_mtu = ifr->ifr_mtu;
838 			PTNET_CORE_UNLOCK(sc);
839 		}
840 		break;
841 
842 	case SIOCSIFMEDIA:
843 	case SIOCGIFMEDIA:
844 		err = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
845 		break;
846 
847 	default:
848 		err = ether_ioctl(ifp, cmd, data);
849 		break;
850 	}
851 
852 	return err;
853 }
854 
855 static int
856 ptnet_init_locked(struct ptnet_softc *sc)
857 {
858 	if_t ifp = sc->ifp;
859 	struct netmap_adapter *na_dr = &sc->ptna->dr.up;
860 	struct netmap_adapter *na_nm = &sc->ptna->hwup.up;
861 	unsigned int nm_buf_size;
862 	int ret;
863 
864 	if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
865 		return 0; /* nothing to do */
866 	}
867 
868 	device_printf(sc->dev, "%s\n", __func__);
869 
870 	/* Translate offload capabilities according to if_capenable. */
871 	ifp->if_hwassist = 0;
872 	if (ifp->if_capenable & IFCAP_TXCSUM)
873 		ifp->if_hwassist |= PTNET_CSUM_OFFLOAD;
874 	if (ifp->if_capenable & IFCAP_TXCSUM_IPV6)
875 		ifp->if_hwassist |= PTNET_CSUM_OFFLOAD_IPV6;
876 	if (ifp->if_capenable & IFCAP_TSO4)
877 		ifp->if_hwassist |= CSUM_IP_TSO;
878 	if (ifp->if_capenable & IFCAP_TSO6)
879 		ifp->if_hwassist |= CSUM_IP6_TSO;
880 
881 	/*
882 	 * Prepare the interface for netmap mode access.
883 	 */
884 	netmap_update_config(na_dr);
885 
886 	ret = netmap_mem_finalize(na_dr->nm_mem, na_dr);
887 	if (ret) {
888 		device_printf(sc->dev, "netmap_mem_finalize() failed\n");
889 		return ret;
890 	}
891 
892 	if (sc->ptna->backend_users == 0) {
893 		ret = ptnet_nm_krings_create(na_nm);
894 		if (ret) {
895 			device_printf(sc->dev, "ptnet_nm_krings_create() "
896 					       "failed\n");
897 			goto err_mem_finalize;
898 		}
899 
900 		ret = netmap_mem_rings_create(na_dr);
901 		if (ret) {
902 			device_printf(sc->dev, "netmap_mem_rings_create() "
903 					       "failed\n");
904 			goto err_rings_create;
905 		}
906 
907 		ret = netmap_mem_get_lut(na_dr->nm_mem, &na_dr->na_lut);
908 		if (ret) {
909 			device_printf(sc->dev, "netmap_mem_get_lut() "
910 					       "failed\n");
911 			goto err_get_lut;
912 		}
913 	}
914 
915 	ret = ptnet_nm_register(na_dr, 1 /* on */);
916 	if (ret) {
917 		goto err_register;
918 	}
919 
920 	nm_buf_size = NETMAP_BUF_SIZE(na_dr);
921 
922 	KASSERT(nm_buf_size > 0, ("Invalid netmap buffer size"));
923 	sc->min_tx_space = PTNET_MAX_PKT_SIZE / nm_buf_size + 2;
924 	device_printf(sc->dev, "%s: min_tx_space = %u\n", __func__,
925 		      sc->min_tx_space);
926 #ifdef PTNETMAP_STATS
927 	callout_reset(&sc->tick, hz, ptnet_tick, sc);
928 #endif
929 
930 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
931 
932 	return 0;
933 
934 err_register:
935 	memset(&na_dr->na_lut, 0, sizeof(na_dr->na_lut));
936 err_get_lut:
937 	netmap_mem_rings_delete(na_dr);
938 err_rings_create:
939 	ptnet_nm_krings_delete(na_nm);
940 err_mem_finalize:
941 	netmap_mem_deref(na_dr->nm_mem, na_dr);
942 
943 	return ret;
944 }
945 
946 /* To be called under core lock. */
947 static int
948 ptnet_stop(struct ptnet_softc *sc)
949 {
950 	if_t ifp = sc->ifp;
951 	struct netmap_adapter *na_dr = &sc->ptna->dr.up;
952 	struct netmap_adapter *na_nm = &sc->ptna->hwup.up;
953 	int i;
954 
955 	device_printf(sc->dev, "%s\n", __func__);
956 
957 	if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) {
958 		return 0; /* nothing to do */
959 	}
960 
961 	/* Clear the driver-ready flag, and synchronize with all the queues,
962 	 * so that after this loop we are sure nobody is working anymore with
963 	 * the device. This scheme is taken from the vtnet driver. */
964 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
965 	callout_stop(&sc->tick);
966 	for (i = 0; i < sc->num_rings; i++) {
967 		PTNET_Q_LOCK(sc->queues + i);
968 		PTNET_Q_UNLOCK(sc->queues + i);
969 	}
970 
971 	ptnet_nm_register(na_dr, 0 /* off */);
972 
973 	if (sc->ptna->backend_users == 0) {
974 		netmap_mem_rings_delete(na_dr);
975 		ptnet_nm_krings_delete(na_nm);
976 	}
977 	netmap_mem_deref(na_dr->nm_mem, na_dr);
978 
979 	return 0;
980 }
981 
982 static void
983 ptnet_qflush(if_t ifp)
984 {
985 	struct ptnet_softc *sc = if_getsoftc(ifp);
986 	int i;
987 
988 	/* Flush all the bufrings and do the interface flush. */
989 	for (i = 0; i < sc->num_rings; i++) {
990 		struct ptnet_queue *pq = sc->queues + i;
991 		struct mbuf *m;
992 
993 		PTNET_Q_LOCK(pq);
994 		if (pq->bufring) {
995 			while ((m = buf_ring_dequeue_sc(pq->bufring))) {
996 				m_freem(m);
997 			}
998 		}
999 		PTNET_Q_UNLOCK(pq);
1000 	}
1001 
1002 	if_qflush(ifp);
1003 }
1004 
1005 static int
1006 ptnet_media_change(if_t ifp)
1007 {
1008 	struct ptnet_softc *sc = if_getsoftc(ifp);
1009 	struct ifmedia *ifm = &sc->media;
1010 
1011 	if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER) {
1012 		return EINVAL;
1013 	}
1014 
1015 	return 0;
1016 }
1017 
1018 #if __FreeBSD_version >= 1100000
1019 static uint64_t
1020 ptnet_get_counter(if_t ifp, ift_counter cnt)
1021 {
1022 	struct ptnet_softc *sc = if_getsoftc(ifp);
1023 	struct ptnet_queue_stats stats[2];
1024 	int i;
1025 
1026 	/* Accumulate statistics over the queues. */
1027 	memset(stats, 0, sizeof(stats));
1028 	for (i = 0; i < sc->num_rings; i++) {
1029 		struct ptnet_queue *pq = sc->queues + i;
1030 		int idx = (i < sc->num_tx_rings) ? 0 : 1;
1031 
1032 		stats[idx].packets	+= pq->stats.packets;
1033 		stats[idx].bytes	+= pq->stats.bytes;
1034 		stats[idx].errors	+= pq->stats.errors;
1035 		stats[idx].iqdrops	+= pq->stats.iqdrops;
1036 		stats[idx].mcasts	+= pq->stats.mcasts;
1037 	}
1038 
1039 	switch (cnt) {
1040 	case IFCOUNTER_IPACKETS:
1041 		return (stats[1].packets);
1042 	case IFCOUNTER_IQDROPS:
1043 		return (stats[1].iqdrops);
1044 	case IFCOUNTER_IERRORS:
1045 		return (stats[1].errors);
1046 	case IFCOUNTER_OPACKETS:
1047 		return (stats[0].packets);
1048 	case IFCOUNTER_OBYTES:
1049 		return (stats[0].bytes);
1050 	case IFCOUNTER_OMCASTS:
1051 		return (stats[0].mcasts);
1052 	default:
1053 		return (if_get_counter_default(ifp, cnt));
1054 	}
1055 }
1056 #endif
1057 
1058 
1059 #ifdef PTNETMAP_STATS
1060 /* Called under core lock. */
1061 static void
1062 ptnet_tick(void *opaque)
1063 {
1064 	struct ptnet_softc *sc = opaque;
1065 	int i;
1066 
1067 	for (i = 0; i < sc->num_rings; i++) {
1068 		struct ptnet_queue *pq = sc->queues + i;
1069 		struct ptnet_queue_stats cur = pq->stats;
1070 		struct timeval now;
1071 		unsigned int delta;
1072 
1073 		microtime(&now);
1074 		delta = now.tv_usec - sc->last_ts.tv_usec +
1075 			(now.tv_sec - sc->last_ts.tv_sec) * 1000000;
1076 		delta /= 1000; /* in milliseconds */
1077 
1078 		if (delta == 0)
1079 			continue;
1080 
1081 		device_printf(sc->dev, "#%d[%u ms]:pkts %lu, kicks %lu, "
1082 			      "intr %lu\n", i, delta,
1083 			      (cur.packets - pq->last_stats.packets),
1084 			      (cur.kicks - pq->last_stats.kicks),
1085 			      (cur.intrs - pq->last_stats.intrs));
1086 		pq->last_stats = cur;
1087 	}
1088 	microtime(&sc->last_ts);
1089 	callout_schedule(&sc->tick, hz);
1090 }
1091 #endif /* PTNETMAP_STATS */
1092 
1093 static void
1094 ptnet_media_status(if_t ifp, struct ifmediareq *ifmr)
1095 {
1096 	/* We are always active, as the backend netmap port is
1097 	 * always open in netmap mode. */
1098 	ifmr->ifm_status = IFM_AVALID | IFM_ACTIVE;
1099 	ifmr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
1100 }
1101 
1102 static uint32_t
1103 ptnet_nm_ptctl(struct ptnet_softc *sc, uint32_t cmd)
1104 {
1105 	/*
1106 	 * Write a command and read back error status,
1107 	 * with zero meaning success.
1108 	 */
1109 	bus_write_4(sc->iomem, PTNET_IO_PTCTL, cmd);
1110 	return bus_read_4(sc->iomem, PTNET_IO_PTCTL);
1111 }
1112 
1113 static int
1114 ptnet_nm_config(struct netmap_adapter *na, struct nm_config_info *info)
1115 {
1116 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1117 
1118 	info->num_tx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_RINGS);
1119 	info->num_rx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_RINGS);
1120 	info->num_tx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_SLOTS);
1121 	info->num_rx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_SLOTS);
1122 	info->rx_buf_maxsize = NETMAP_BUF_SIZE(na);
1123 
1124 	device_printf(sc->dev, "txr %u, rxr %u, txd %u, rxd %u, rxbufsz %u\n",
1125 			info->num_tx_rings, info->num_rx_rings,
1126 			info->num_tx_descs, info->num_rx_descs,
1127 			info->rx_buf_maxsize);
1128 
1129 	return 0;
1130 }
1131 
1132 static void
1133 ptnet_sync_from_csb(struct ptnet_softc *sc, struct netmap_adapter *na)
1134 {
1135 	int i;
1136 
1137 	/* Sync krings from the host, reading from
1138 	 * CSB. */
1139 	for (i = 0; i < sc->num_rings; i++) {
1140 		struct nm_csb_atok *atok = sc->queues[i].atok;
1141 		struct nm_csb_ktoa *ktoa = sc->queues[i].ktoa;
1142 		struct netmap_kring *kring;
1143 
1144 		if (i < na->num_tx_rings) {
1145 			kring = na->tx_rings[i];
1146 		} else {
1147 			kring = na->rx_rings[i - na->num_tx_rings];
1148 		}
1149 		kring->rhead = kring->ring->head = atok->head;
1150 		kring->rcur = kring->ring->cur = atok->cur;
1151 		kring->nr_hwcur = ktoa->hwcur;
1152 		kring->nr_hwtail = kring->rtail =
1153 			kring->ring->tail = ktoa->hwtail;
1154 
1155 		nm_prdis("%d,%d: csb {hc %u h %u c %u ht %u}", t, i,
1156 		   ktoa->hwcur, atok->head, atok->cur,
1157 		   ktoa->hwtail);
1158 		nm_prdis("%d,%d: kring {hc %u rh %u rc %u h %u c %u ht %u rt %u t %u}",
1159 		   t, i, kring->nr_hwcur, kring->rhead, kring->rcur,
1160 		   kring->ring->head, kring->ring->cur, kring->nr_hwtail,
1161 		   kring->rtail, kring->ring->tail);
1162 	}
1163 }
1164 
1165 static void
1166 ptnet_update_vnet_hdr(struct ptnet_softc *sc)
1167 {
1168 	unsigned int wanted_hdr_len = ptnet_vnet_hdr ? PTNET_HDR_SIZE : 0;
1169 
1170 	bus_write_4(sc->iomem, PTNET_IO_VNET_HDR_LEN, wanted_hdr_len);
1171 	sc->vnet_hdr_len = bus_read_4(sc->iomem, PTNET_IO_VNET_HDR_LEN);
1172 	sc->ptna->hwup.up.virt_hdr_len = sc->vnet_hdr_len;
1173 }
1174 
1175 static int
1176 ptnet_nm_register(struct netmap_adapter *na, int onoff)
1177 {
1178 	/* device-specific */
1179 	if_t ifp = na->ifp;
1180 	struct ptnet_softc *sc = if_getsoftc(ifp);
1181 	int native = (na == &sc->ptna->hwup.up);
1182 	struct ptnet_queue *pq;
1183 	int ret = 0;
1184 	int i;
1185 
1186 	if (!onoff) {
1187 		sc->ptna->backend_users--;
1188 	}
1189 
1190 	/* If this is the last netmap client, guest interrupt enable flags may
1191 	 * be in arbitrary state. Since these flags are going to be used also
1192 	 * by the netdevice driver, we have to make sure to start with
1193 	 * notifications enabled. Also, schedule NAPI to flush pending packets
1194 	 * in the RX rings, since we will not receive further interrupts
1195 	 * until these will be processed. */
1196 	if (native && !onoff && na->active_fds == 0) {
1197 		nm_prinf("Exit netmap mode, re-enable interrupts");
1198 		for (i = 0; i < sc->num_rings; i++) {
1199 			pq = sc->queues + i;
1200 			pq->atok->appl_need_kick = 1;
1201 		}
1202 	}
1203 
1204 	if (onoff) {
1205 		if (sc->ptna->backend_users == 0) {
1206 			/* Initialize notification enable fields in the CSB. */
1207 			for (i = 0; i < sc->num_rings; i++) {
1208 				pq = sc->queues + i;
1209 				pq->ktoa->kern_need_kick = 1;
1210 				pq->atok->appl_need_kick =
1211 					(!(ifp->if_capenable & IFCAP_POLLING)
1212 						&& i >= sc->num_tx_rings);
1213 			}
1214 
1215 			/* Set the virtio-net header length. */
1216 			ptnet_update_vnet_hdr(sc);
1217 
1218 			/* Make sure the host adapter passed through is ready
1219 			 * for txsync/rxsync. */
1220 			ret = ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_CREATE);
1221 			if (ret) {
1222 				return ret;
1223 			}
1224 
1225 			/* Align the guest krings and rings to the state stored
1226 			 * in the CSB. */
1227 			ptnet_sync_from_csb(sc, na);
1228 		}
1229 
1230 		/* If not native, don't call nm_set_native_flags, since we don't want
1231 		 * to replace if_transmit method, nor set NAF_NETMAP_ON */
1232 		if (native) {
1233 			netmap_krings_mode_commit(na, onoff);
1234 			nm_set_native_flags(na);
1235 		}
1236 
1237 	} else {
1238 		if (native) {
1239 			nm_clear_native_flags(na);
1240 			netmap_krings_mode_commit(na, onoff);
1241 		}
1242 
1243 		if (sc->ptna->backend_users == 0) {
1244 			ret = ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_DELETE);
1245 		}
1246 	}
1247 
1248 	if (onoff) {
1249 		sc->ptna->backend_users++;
1250 	}
1251 
1252 	return ret;
1253 }
1254 
1255 static int
1256 ptnet_nm_txsync(struct netmap_kring *kring, int flags)
1257 {
1258 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1259 	struct ptnet_queue *pq = sc->queues + kring->ring_id;
1260 	bool notify;
1261 
1262 	notify = netmap_pt_guest_txsync(pq->atok, pq->ktoa, kring, flags);
1263 	if (notify) {
1264 		ptnet_kick(pq);
1265 	}
1266 
1267 	return 0;
1268 }
1269 
1270 static int
1271 ptnet_nm_rxsync(struct netmap_kring *kring, int flags)
1272 {
1273 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1274 	struct ptnet_queue *pq = sc->rxqueues + kring->ring_id;
1275 	bool notify;
1276 
1277 	notify = netmap_pt_guest_rxsync(pq->atok, pq->ktoa, kring, flags);
1278 	if (notify) {
1279 		ptnet_kick(pq);
1280 	}
1281 
1282 	return 0;
1283 }
1284 
1285 static void
1286 ptnet_nm_intr(struct netmap_adapter *na, int onoff)
1287 {
1288 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1289 	int i;
1290 
1291 	for (i = 0; i < sc->num_rings; i++) {
1292 		struct ptnet_queue *pq = sc->queues + i;
1293 		pq->atok->appl_need_kick = onoff;
1294 	}
1295 }
1296 
1297 static void
1298 ptnet_tx_intr(void *opaque)
1299 {
1300 	struct ptnet_queue *pq = opaque;
1301 	struct ptnet_softc *sc = pq->sc;
1302 
1303 	DBG(device_printf(sc->dev, "Tx interrupt #%d\n", pq->kring_id));
1304 #ifdef PTNETMAP_STATS
1305 	pq->stats.intrs ++;
1306 #endif /* PTNETMAP_STATS */
1307 
1308 	if (netmap_tx_irq(sc->ifp, pq->kring_id) != NM_IRQ_PASS) {
1309 		return;
1310 	}
1311 
1312 	/* Schedule the tasqueue to flush process transmissions requests.
1313 	 * However, vtnet, if_em and if_igb just call ptnet_transmit() here,
1314 	 * at least when using MSI-X interrupts. The if_em driver, instead
1315 	 * schedule taskqueue when using legacy interrupts. */
1316 	taskqueue_enqueue(pq->taskq, &pq->task);
1317 }
1318 
1319 static void
1320 ptnet_rx_intr(void *opaque)
1321 {
1322 	struct ptnet_queue *pq = opaque;
1323 	struct ptnet_softc *sc = pq->sc;
1324 	unsigned int unused;
1325 
1326 	DBG(device_printf(sc->dev, "Rx interrupt #%d\n", pq->kring_id));
1327 #ifdef PTNETMAP_STATS
1328 	pq->stats.intrs ++;
1329 #endif /* PTNETMAP_STATS */
1330 
1331 	if (netmap_rx_irq(sc->ifp, pq->kring_id, &unused) != NM_IRQ_PASS) {
1332 		return;
1333 	}
1334 
1335 	/* Like vtnet, if_igb and if_em drivers when using MSI-X interrupts,
1336 	 * receive-side processing is executed directly in the interrupt
1337 	 * service routine. Alternatively, we may schedule the taskqueue. */
1338 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
1339 }
1340 
1341 static void
1342 ptnet_vlan_tag_remove(struct mbuf *m)
1343 {
1344 	struct ether_vlan_header *evh;
1345 
1346 	evh = mtod(m, struct ether_vlan_header *);
1347 	m->m_pkthdr.ether_vtag = ntohs(evh->evl_tag);
1348 	m->m_flags |= M_VLANTAG;
1349 
1350 	/* Strip the 802.1Q header. */
1351 	bcopy((char *) evh, (char *) evh + ETHER_VLAN_ENCAP_LEN,
1352 	    ETHER_HDR_LEN - ETHER_TYPE_LEN);
1353 	m_adj(m, ETHER_VLAN_ENCAP_LEN);
1354 }
1355 
1356 static void
1357 ptnet_ring_update(struct ptnet_queue *pq, struct netmap_kring *kring,
1358 		  unsigned int head, unsigned int sync_flags)
1359 {
1360 	struct netmap_ring *ring = kring->ring;
1361 	struct nm_csb_atok *atok = pq->atok;
1362 	struct nm_csb_ktoa *ktoa = pq->ktoa;
1363 
1364 	/* Some packets have been pushed to the netmap ring. We have
1365 	 * to tell the host to process the new packets, updating cur
1366 	 * and head in the CSB. */
1367 	ring->head = ring->cur = head;
1368 
1369 	/* Mimic nm_txsync_prologue/nm_rxsync_prologue. */
1370 	kring->rcur = kring->rhead = head;
1371 
1372 	nm_sync_kloop_appl_write(atok, kring->rcur, kring->rhead);
1373 
1374 	/* Kick the host if needed. */
1375 	if (NM_ACCESS_ONCE(ktoa->kern_need_kick)) {
1376 		atok->sync_flags = sync_flags;
1377 		ptnet_kick(pq);
1378 	}
1379 }
1380 
1381 #define PTNET_TX_NOSPACE(_h, _k, _min)	\
1382 	((((_h) < (_k)->rtail) ? 0 : (_k)->nkr_num_slots) + \
1383 		(_k)->rtail - (_h)) < (_min)
1384 
1385 /* This function may be called by the network stack, or by
1386  * by the taskqueue thread. */
1387 static int
1388 ptnet_drain_transmit_queue(struct ptnet_queue *pq, unsigned int budget,
1389 			   bool may_resched)
1390 {
1391 	struct ptnet_softc *sc = pq->sc;
1392 	bool have_vnet_hdr = sc->vnet_hdr_len;
1393 	struct netmap_adapter *na = &sc->ptna->dr.up;
1394 	if_t ifp = sc->ifp;
1395 	unsigned int batch_count = 0;
1396 	struct nm_csb_atok *atok;
1397 	struct nm_csb_ktoa *ktoa;
1398 	struct netmap_kring *kring;
1399 	struct netmap_ring *ring;
1400 	struct netmap_slot *slot;
1401 	unsigned int count = 0;
1402 	unsigned int minspace;
1403 	unsigned int head;
1404 	unsigned int lim;
1405 	struct mbuf *mhead;
1406 	struct mbuf *mf;
1407 	int nmbuf_bytes;
1408 	uint8_t *nmbuf;
1409 
1410 	if (!PTNET_Q_TRYLOCK(pq)) {
1411 		/* We failed to acquire the lock, schedule the taskqueue. */
1412 		nm_prlim(1, "Deferring TX work");
1413 		if (may_resched) {
1414 			taskqueue_enqueue(pq->taskq, &pq->task);
1415 		}
1416 
1417 		return 0;
1418 	}
1419 
1420 	if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1421 		PTNET_Q_UNLOCK(pq);
1422 		nm_prlim(1, "Interface is down");
1423 		return ENETDOWN;
1424 	}
1425 
1426 	atok = pq->atok;
1427 	ktoa = pq->ktoa;
1428 	kring = na->tx_rings[pq->kring_id];
1429 	ring = kring->ring;
1430 	lim = kring->nkr_num_slots - 1;
1431 	head = ring->head;
1432 	minspace = sc->min_tx_space;
1433 
1434 	while (count < budget) {
1435 		if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1436 			/* We ran out of slot, let's see if the host has
1437 			 * freed up some, by reading hwcur and hwtail from
1438 			 * the CSB. */
1439 			ptnet_sync_tail(ktoa, kring);
1440 
1441 			if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1442 				/* Still no slots available. Reactivate the
1443 				 * interrupts so that we can be notified
1444 				 * when some free slots are made available by
1445 				 * the host. */
1446 				atok->appl_need_kick = 1;
1447 
1448 				/* Double check. We need a full barrier to
1449 				 * prevent the store to atok->appl_need_kick
1450 				 * to be reordered with the load from
1451 				 * ktoa->hwcur and ktoa->hwtail (store-load
1452 				 * barrier). */
1453 				nm_stld_barrier();
1454 				ptnet_sync_tail(ktoa, kring);
1455 				if (likely(PTNET_TX_NOSPACE(head, kring,
1456 							    minspace))) {
1457 					break;
1458 				}
1459 
1460 				nm_prlim(1, "Found more slots by doublecheck");
1461 				/* More slots were freed before reactivating
1462 				 * the interrupts. */
1463 				atok->appl_need_kick = 0;
1464 			}
1465 		}
1466 
1467 		mhead = drbr_peek(ifp, pq->bufring);
1468 		if (!mhead) {
1469 			break;
1470 		}
1471 
1472 		/* Initialize transmission state variables. */
1473 		slot = ring->slot + head;
1474 		nmbuf = NMB(na, slot);
1475 		nmbuf_bytes = 0;
1476 
1477 		/* If needed, prepare the virtio-net header at the beginning
1478 		 * of the first slot. */
1479 		if (have_vnet_hdr) {
1480 			struct virtio_net_hdr *vh =
1481 					(struct virtio_net_hdr *)nmbuf;
1482 
1483 			/* For performance, we could replace this memset() with
1484 			 * two 8-bytes-wide writes. */
1485 			memset(nmbuf, 0, PTNET_HDR_SIZE);
1486 			if (mhead->m_pkthdr.csum_flags & PTNET_ALL_OFFLOAD) {
1487 				mhead = virtio_net_tx_offload(ifp, mhead, false,
1488 							 vh);
1489 				if (unlikely(!mhead)) {
1490 					/* Packet dropped because errors
1491 					 * occurred while preparing the vnet
1492 					 * header. Let's go ahead with the next
1493 					 * packet. */
1494 					pq->stats.errors ++;
1495 					drbr_advance(ifp, pq->bufring);
1496 					continue;
1497 				}
1498 			}
1499 			nm_prdis(1, "%s: [csum_flags %lX] vnet hdr: flags %x "
1500 			      "csum_start %u csum_ofs %u hdr_len = %u "
1501 			      "gso_size %u gso_type %x", __func__,
1502 			      mhead->m_pkthdr.csum_flags, vh->flags,
1503 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
1504 			      vh->gso_size, vh->gso_type);
1505 
1506 			nmbuf += PTNET_HDR_SIZE;
1507 			nmbuf_bytes += PTNET_HDR_SIZE;
1508 		}
1509 
1510 		for (mf = mhead; mf; mf = mf->m_next) {
1511 			uint8_t *mdata = mf->m_data;
1512 			int mlen = mf->m_len;
1513 
1514 			for (;;) {
1515 				int copy = NETMAP_BUF_SIZE(na) - nmbuf_bytes;
1516 
1517 				if (mlen < copy) {
1518 					copy = mlen;
1519 				}
1520 				memcpy(nmbuf, mdata, copy);
1521 
1522 				mdata += copy;
1523 				mlen -= copy;
1524 				nmbuf += copy;
1525 				nmbuf_bytes += copy;
1526 
1527 				if (!mlen) {
1528 					break;
1529 				}
1530 
1531 				slot->len = nmbuf_bytes;
1532 				slot->flags = NS_MOREFRAG;
1533 
1534 				head = nm_next(head, lim);
1535 				KASSERT(head != ring->tail,
1536 					("Unexpectedly run out of TX space"));
1537 				slot = ring->slot + head;
1538 				nmbuf = NMB(na, slot);
1539 				nmbuf_bytes = 0;
1540 			}
1541 		}
1542 
1543 		/* Complete last slot and update head. */
1544 		slot->len = nmbuf_bytes;
1545 		slot->flags = 0;
1546 		head = nm_next(head, lim);
1547 
1548 		/* Consume the packet just processed. */
1549 		drbr_advance(ifp, pq->bufring);
1550 
1551 		/* Copy the packet to listeners. */
1552 		ETHER_BPF_MTAP(ifp, mhead);
1553 
1554 		pq->stats.packets ++;
1555 		pq->stats.bytes += mhead->m_pkthdr.len;
1556 		if (mhead->m_flags & M_MCAST) {
1557 			pq->stats.mcasts ++;
1558 		}
1559 
1560 		m_freem(mhead);
1561 
1562 		count ++;
1563 		if (++batch_count == PTNET_TX_BATCH) {
1564 			ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1565 			batch_count = 0;
1566 		}
1567 	}
1568 
1569 	if (batch_count) {
1570 		ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1571 	}
1572 
1573 	if (count >= budget && may_resched) {
1574 		DBG(nm_prlim(1, "out of budget: resched, %d mbufs pending\n",
1575 					drbr_inuse(ifp, pq->bufring)));
1576 		taskqueue_enqueue(pq->taskq, &pq->task);
1577 	}
1578 
1579 	PTNET_Q_UNLOCK(pq);
1580 
1581 	return count;
1582 }
1583 
1584 static int
1585 ptnet_transmit(if_t ifp, struct mbuf *m)
1586 {
1587 	struct ptnet_softc *sc = if_getsoftc(ifp);
1588 	struct ptnet_queue *pq;
1589 	unsigned int queue_idx;
1590 	int err;
1591 
1592 	DBG(device_printf(sc->dev, "transmit %p\n", m));
1593 
1594 	/* Insert 802.1Q header if needed. */
1595 	if (m->m_flags & M_VLANTAG) {
1596 		m = ether_vlanencap(m, m->m_pkthdr.ether_vtag);
1597 		if (m == NULL) {
1598 			return ENOBUFS;
1599 		}
1600 		m->m_flags &= ~M_VLANTAG;
1601 	}
1602 
1603 	/* Get the flow-id if available. */
1604 	queue_idx = (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) ?
1605 		    m->m_pkthdr.flowid : curcpu;
1606 
1607 	if (unlikely(queue_idx >= sc->num_tx_rings)) {
1608 		queue_idx %= sc->num_tx_rings;
1609 	}
1610 
1611 	pq = sc->queues + queue_idx;
1612 
1613 	err = drbr_enqueue(ifp, pq->bufring, m);
1614 	if (err) {
1615 		/* ENOBUFS when the bufring is full */
1616 		nm_prlim(1, "%s: drbr_enqueue() failed %d\n",
1617 			__func__, err);
1618 		pq->stats.errors ++;
1619 		return err;
1620 	}
1621 
1622 	if (ifp->if_capenable & IFCAP_POLLING) {
1623 		/* If polling is on, the transmit queues will be
1624 		 * drained by the poller. */
1625 		return 0;
1626 	}
1627 
1628 	err = ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
1629 
1630 	return (err < 0) ? err : 0;
1631 }
1632 
1633 static unsigned int
1634 ptnet_rx_discard(struct netmap_kring *kring, unsigned int head)
1635 {
1636 	struct netmap_ring *ring = kring->ring;
1637 	struct netmap_slot *slot = ring->slot + head;
1638 
1639 	for (;;) {
1640 		head = nm_next(head, kring->nkr_num_slots - 1);
1641 		if (!(slot->flags & NS_MOREFRAG) || head == ring->tail) {
1642 			break;
1643 		}
1644 		slot = ring->slot + head;
1645 	}
1646 
1647 	return head;
1648 }
1649 
1650 static inline struct mbuf *
1651 ptnet_rx_slot(struct mbuf *mtail, uint8_t *nmbuf, unsigned int nmbuf_len)
1652 {
1653 	uint8_t *mdata = mtod(mtail, uint8_t *) + mtail->m_len;
1654 
1655 	do {
1656 		unsigned int copy;
1657 
1658 		if (mtail->m_len == MCLBYTES) {
1659 			struct mbuf *mf;
1660 
1661 			mf = m_getcl(M_NOWAIT, MT_DATA, 0);
1662 			if (unlikely(!mf)) {
1663 				return NULL;
1664 			}
1665 
1666 			mtail->m_next = mf;
1667 			mtail = mf;
1668 			mdata = mtod(mtail, uint8_t *);
1669 			mtail->m_len = 0;
1670 		}
1671 
1672 		copy = MCLBYTES - mtail->m_len;
1673 		if (nmbuf_len < copy) {
1674 			copy = nmbuf_len;
1675 		}
1676 
1677 		memcpy(mdata, nmbuf, copy);
1678 
1679 		nmbuf += copy;
1680 		nmbuf_len -= copy;
1681 		mdata += copy;
1682 		mtail->m_len += copy;
1683 	} while (nmbuf_len);
1684 
1685 	return mtail;
1686 }
1687 
1688 static int
1689 ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget, bool may_resched)
1690 {
1691 	struct ptnet_softc *sc = pq->sc;
1692 	bool have_vnet_hdr = sc->vnet_hdr_len;
1693 	struct nm_csb_atok *atok = pq->atok;
1694 	struct nm_csb_ktoa *ktoa = pq->ktoa;
1695 	struct netmap_adapter *na = &sc->ptna->dr.up;
1696 	struct netmap_kring *kring = na->rx_rings[pq->kring_id];
1697 	struct netmap_ring *ring = kring->ring;
1698 	unsigned int const lim = kring->nkr_num_slots - 1;
1699 	unsigned int batch_count = 0;
1700 	if_t ifp = sc->ifp;
1701 	unsigned int count = 0;
1702 	uint32_t head;
1703 
1704 	PTNET_Q_LOCK(pq);
1705 
1706 	if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1707 		goto unlock;
1708 	}
1709 
1710 	kring->nr_kflags &= ~NKR_PENDINTR;
1711 
1712 	head = ring->head;
1713 	while (count < budget) {
1714 		uint32_t prev_head = head;
1715 		struct mbuf *mhead, *mtail;
1716 		struct virtio_net_hdr *vh;
1717 		struct netmap_slot *slot;
1718 		unsigned int nmbuf_len;
1719 		uint8_t *nmbuf;
1720 		int deliver = 1; /* the mbuf to the network stack. */
1721 host_sync:
1722 		if (head == ring->tail) {
1723 			/* We ran out of slot, let's see if the host has
1724 			 * added some, by reading hwcur and hwtail from
1725 			 * the CSB. */
1726 			ptnet_sync_tail(ktoa, kring);
1727 
1728 			if (head == ring->tail) {
1729 				/* Still no slots available. Reactivate
1730 				 * interrupts as they were disabled by the
1731 				 * host thread right before issuing the
1732 				 * last interrupt. */
1733 				atok->appl_need_kick = 1;
1734 
1735 				/* Double check for more completed RX slots.
1736 				 * We need a full barrier to prevent the store
1737 				 * to atok->appl_need_kick to be reordered with
1738 				 * the load from ktoa->hwcur and ktoa->hwtail
1739 				 * (store-load barrier). */
1740 				nm_stld_barrier();
1741 				ptnet_sync_tail(ktoa, kring);
1742 				if (likely(head == ring->tail)) {
1743 					break;
1744 				}
1745 				atok->appl_need_kick = 0;
1746 			}
1747 		}
1748 
1749 		/* Initialize ring state variables, possibly grabbing the
1750 		 * virtio-net header. */
1751 		slot = ring->slot + head;
1752 		nmbuf = NMB(na, slot);
1753 		nmbuf_len = slot->len;
1754 
1755 		vh = (struct virtio_net_hdr *)nmbuf;
1756 		if (have_vnet_hdr) {
1757 			if (unlikely(nmbuf_len < PTNET_HDR_SIZE)) {
1758 				/* There is no good reason why host should
1759 				 * put the header in multiple netmap slots.
1760 				 * If this is the case, discard. */
1761 				nm_prlim(1, "Fragmented vnet-hdr: dropping");
1762 				head = ptnet_rx_discard(kring, head);
1763 				pq->stats.iqdrops ++;
1764 				deliver = 0;
1765 				goto skip;
1766 			}
1767 			nm_prdis(1, "%s: vnet hdr: flags %x csum_start %u "
1768 			      "csum_ofs %u hdr_len = %u gso_size %u "
1769 			      "gso_type %x", __func__, vh->flags,
1770 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
1771 			      vh->gso_size, vh->gso_type);
1772 			nmbuf += PTNET_HDR_SIZE;
1773 			nmbuf_len -= PTNET_HDR_SIZE;
1774 		}
1775 
1776 		/* Allocate the head of a new mbuf chain.
1777 		 * We use m_getcl() to allocate an mbuf with standard cluster
1778 		 * size (MCLBYTES). In the future we could use m_getjcl()
1779 		 * to choose different sizes. */
1780 		mhead = mtail = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1781 		if (unlikely(mhead == NULL)) {
1782 			device_printf(sc->dev, "%s: failed to allocate mbuf "
1783 				      "head\n", __func__);
1784 			pq->stats.errors ++;
1785 			break;
1786 		}
1787 
1788 		/* Initialize the mbuf state variables. */
1789 		mhead->m_pkthdr.len = nmbuf_len;
1790 		mtail->m_len = 0;
1791 
1792 		/* Scan all the netmap slots containing the current packet. */
1793 		for (;;) {
1794 			DBG(device_printf(sc->dev, "%s: h %u t %u rcv frag "
1795 					  "len %u, flags %u\n", __func__,
1796 					  head, ring->tail, slot->len,
1797 					  slot->flags));
1798 
1799 			mtail = ptnet_rx_slot(mtail, nmbuf, nmbuf_len);
1800 			if (unlikely(!mtail)) {
1801 				/* Ouch. We ran out of memory while processing
1802 				 * a packet. We have to restore the previous
1803 				 * head position, free the mbuf chain, and
1804 				 * schedule the taskqueue to give the packet
1805 				 * another chance. */
1806 				device_printf(sc->dev, "%s: failed to allocate"
1807 					" mbuf frag, reset head %u --> %u\n",
1808 					__func__, head, prev_head);
1809 				head = prev_head;
1810 				m_freem(mhead);
1811 				pq->stats.errors ++;
1812 				if (may_resched) {
1813 					taskqueue_enqueue(pq->taskq,
1814 							  &pq->task);
1815 				}
1816 				goto escape;
1817 			}
1818 
1819 			/* We have to increment head irrespective of the
1820 			 * NS_MOREFRAG being set or not. */
1821 			head = nm_next(head, lim);
1822 
1823 			if (!(slot->flags & NS_MOREFRAG)) {
1824 				break;
1825 			}
1826 
1827 			if (unlikely(head == ring->tail)) {
1828 				/* The very last slot prepared by the host has
1829 				 * the NS_MOREFRAG set. Drop it and continue
1830 				 * the outer cycle (to do the double-check). */
1831 				nm_prlim(1, "Incomplete packet: dropping");
1832 				m_freem(mhead);
1833 				pq->stats.iqdrops ++;
1834 				goto host_sync;
1835 			}
1836 
1837 			slot = ring->slot + head;
1838 			nmbuf = NMB(na, slot);
1839 			nmbuf_len = slot->len;
1840 			mhead->m_pkthdr.len += nmbuf_len;
1841 		}
1842 
1843 		mhead->m_pkthdr.rcvif = ifp;
1844 		mhead->m_pkthdr.csum_flags = 0;
1845 
1846 		/* Store the queue idx in the packet header. */
1847 		mhead->m_pkthdr.flowid = pq->kring_id;
1848 		M_HASHTYPE_SET(mhead, M_HASHTYPE_OPAQUE);
1849 
1850 		if (ifp->if_capenable & IFCAP_VLAN_HWTAGGING) {
1851 			struct ether_header *eh;
1852 
1853 			eh = mtod(mhead, struct ether_header *);
1854 			if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
1855 				ptnet_vlan_tag_remove(mhead);
1856 				/*
1857 				 * With the 802.1Q header removed, update the
1858 				 * checksum starting location accordingly.
1859 				 */
1860 				if (vh->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
1861 					vh->csum_start -= ETHER_VLAN_ENCAP_LEN;
1862 			}
1863 		}
1864 
1865 		if (unlikely(have_vnet_hdr && virtio_net_rx_csum(mhead, vh))) {
1866 			m_freem(mhead);
1867 			nm_prlim(1, "Csum offload error: dropping");
1868 			pq->stats.iqdrops ++;
1869 			deliver = 0;
1870 		}
1871 
1872 skip:
1873 		count ++;
1874 		if (++batch_count >= PTNET_RX_BATCH) {
1875 			/* Some packets have been (or will be) pushed to the network
1876 			 * stack. We need to update the CSB to tell the host about
1877 			 * the new ring->cur and ring->head (RX buffer refill). */
1878 			ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
1879 			batch_count = 0;
1880 		}
1881 
1882 		if (likely(deliver))  {
1883 			pq->stats.packets ++;
1884 			pq->stats.bytes += mhead->m_pkthdr.len;
1885 
1886 			PTNET_Q_UNLOCK(pq);
1887 			(*ifp->if_input)(ifp, mhead);
1888 			PTNET_Q_LOCK(pq);
1889 			/* The ring->head index (and related indices) are
1890 			 * updated under pq lock by ptnet_ring_update().
1891 			 * Since we dropped the lock to call if_input(), we
1892 			 * must reload ring->head and restart processing the
1893 			 * ring from there. */
1894 			head = ring->head;
1895 
1896 			if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1897 				/* The interface has gone down while we didn't
1898 				 * have the lock. Stop any processing and exit. */
1899 				goto unlock;
1900 			}
1901 		}
1902 	}
1903 escape:
1904 	if (batch_count) {
1905 		ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
1906 
1907 	}
1908 
1909 	if (count >= budget && may_resched) {
1910 		/* If we ran out of budget or the double-check found new
1911 		 * slots to process, schedule the taskqueue. */
1912 		DBG(nm_prlim(1, "out of budget: resched h %u t %u\n",
1913 					head, ring->tail));
1914 		taskqueue_enqueue(pq->taskq, &pq->task);
1915 	}
1916 unlock:
1917 	PTNET_Q_UNLOCK(pq);
1918 
1919 	return count;
1920 }
1921 
1922 static void
1923 ptnet_rx_task(void *context, int pending)
1924 {
1925 	struct ptnet_queue *pq = context;
1926 
1927 	DBG(nm_prlim(1, "%s: pq #%u\n", __func__, pq->kring_id));
1928 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
1929 }
1930 
1931 static void
1932 ptnet_tx_task(void *context, int pending)
1933 {
1934 	struct ptnet_queue *pq = context;
1935 
1936 	DBG(nm_prlim(1, "%s: pq #%u\n", __func__, pq->kring_id));
1937 	ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
1938 }
1939 
1940 #ifdef DEVICE_POLLING
1941 /* We don't need to handle differently POLL_AND_CHECK_STATUS and
1942  * POLL_ONLY, since we don't have an Interrupt Status Register. */
1943 static int
1944 ptnet_poll(if_t ifp, enum poll_cmd cmd, int budget)
1945 {
1946 	struct ptnet_softc *sc = if_getsoftc(ifp);
1947 	unsigned int queue_budget;
1948 	unsigned int count = 0;
1949 	bool borrow = false;
1950 	int i;
1951 
1952 	KASSERT(sc->num_rings > 0, ("Found no queues in while polling ptnet"));
1953 	queue_budget = MAX(budget / sc->num_rings, 1);
1954 	nm_prlim(1, "Per-queue budget is %d", queue_budget);
1955 
1956 	while (budget) {
1957 		unsigned int rcnt = 0;
1958 
1959 		for (i = 0; i < sc->num_rings; i++) {
1960 			struct ptnet_queue *pq = sc->queues + i;
1961 
1962 			if (borrow) {
1963 				queue_budget = MIN(queue_budget, budget);
1964 				if (queue_budget == 0) {
1965 					break;
1966 				}
1967 			}
1968 
1969 			if (i < sc->num_tx_rings) {
1970 				rcnt += ptnet_drain_transmit_queue(pq,
1971 						   queue_budget, false);
1972 			} else {
1973 				rcnt += ptnet_rx_eof(pq, queue_budget,
1974 						      false);
1975 			}
1976 		}
1977 
1978 		if (!rcnt) {
1979 			/* A scan of the queues gave no result, we can
1980 			 * stop here. */
1981 			break;
1982 		}
1983 
1984 		if (rcnt > budget) {
1985 			/* This may happen when initial budget < sc->num_rings,
1986 			 * since one packet budget is given to each queue
1987 			 * anyway. Just pretend we didn't eat "so much". */
1988 			rcnt = budget;
1989 		}
1990 		count += rcnt;
1991 		budget -= rcnt;
1992 		borrow = true;
1993 	}
1994 
1995 
1996 	return count;
1997 }
1998 #endif /* DEVICE_POLLING */
1999 #endif /* WITH_PTNETMAP */
2000