xref: /freebsd/sys/dev/netmap/if_ptnet.c (revision 655d043b496e19e4e1463a12278a895d0d40b33d)
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 static uint64_t	ptnet_get_counter(if_t, ift_counter);
97 
98 //#define PTNETMAP_STATS
99 //#define DEBUG
100 #ifdef DEBUG
101 #define DBG(x) x
102 #else   /* !DEBUG */
103 #define DBG(x)
104 #endif  /* !DEBUG */
105 
106 extern int ptnet_vnet_hdr; /* Tunable parameter */
107 
108 struct ptnet_softc;
109 
110 struct ptnet_queue_stats {
111 	uint64_t	packets; /* if_[io]packets */
112 	uint64_t	bytes;	 /* if_[io]bytes */
113 	uint64_t	errors;	 /* if_[io]errors */
114 	uint64_t	iqdrops; /* if_iqdrops */
115 	uint64_t	mcasts;  /* if_[io]mcasts */
116 #ifdef PTNETMAP_STATS
117 	uint64_t	intrs;
118 	uint64_t	kicks;
119 #endif /* PTNETMAP_STATS */
120 };
121 
122 struct ptnet_queue {
123 	struct ptnet_softc		*sc;
124 	struct				resource *irq;
125 	void				*cookie;
126 	int				kring_id;
127 	struct nm_csb_atok		*atok;
128 	struct nm_csb_ktoa		*ktoa;
129 	unsigned int			kick;
130 	struct mtx			lock;
131 	struct buf_ring			*bufring; /* for TX queues */
132 	struct ptnet_queue_stats	stats;
133 #ifdef PTNETMAP_STATS
134 	struct ptnet_queue_stats	last_stats;
135 #endif /* PTNETMAP_STATS */
136 	struct taskqueue		*taskq;
137 	struct task			task;
138 	char				lock_name[16];
139 };
140 
141 #define PTNET_Q_LOCK(_pq)	mtx_lock(&(_pq)->lock)
142 #define PTNET_Q_TRYLOCK(_pq)	mtx_trylock(&(_pq)->lock)
143 #define PTNET_Q_UNLOCK(_pq)	mtx_unlock(&(_pq)->lock)
144 
145 struct ptnet_softc {
146 	device_t		dev;
147 	if_t			ifp;
148 	struct ifmedia		media;
149 	struct mtx		lock;
150 	char			lock_name[16];
151 	char			hwaddr[ETHER_ADDR_LEN];
152 
153 	/* Mirror of PTFEAT register. */
154 	uint32_t		ptfeatures;
155 	unsigned int		vnet_hdr_len;
156 
157 	/* PCI BARs support. */
158 	struct resource		*iomem;
159 	struct resource		*msix_mem;
160 
161 	unsigned int		num_rings;
162 	unsigned int		num_tx_rings;
163 	struct ptnet_queue	*queues;
164 	struct ptnet_queue	*rxqueues;
165 	struct nm_csb_atok	*csb_gh;
166 	struct nm_csb_ktoa	*csb_hg;
167 
168 	unsigned int		min_tx_space;
169 
170 	struct netmap_pt_guest_adapter *ptna;
171 
172 	struct callout		tick;
173 #ifdef PTNETMAP_STATS
174 	struct timeval		last_ts;
175 #endif /* PTNETMAP_STATS */
176 };
177 
178 #define PTNET_CORE_LOCK(_sc)	mtx_lock(&(_sc)->lock)
179 #define PTNET_CORE_UNLOCK(_sc)	mtx_unlock(&(_sc)->lock)
180 
181 static int	ptnet_probe(device_t);
182 static int	ptnet_attach(device_t);
183 static int	ptnet_detach(device_t);
184 static int	ptnet_suspend(device_t);
185 static int	ptnet_resume(device_t);
186 static int	ptnet_shutdown(device_t);
187 
188 static void	ptnet_init(void *opaque);
189 static int	ptnet_ioctl(if_t ifp, u_long cmd, caddr_t data);
190 static int	ptnet_init_locked(struct ptnet_softc *sc);
191 static int	ptnet_stop(struct ptnet_softc *sc);
192 static int	ptnet_transmit(if_t ifp, struct mbuf *m);
193 static int	ptnet_drain_transmit_queue(struct ptnet_queue *pq,
194 					   unsigned int budget,
195 					   bool may_resched);
196 static void	ptnet_qflush(if_t ifp);
197 static void	ptnet_tx_task(void *context, int pending);
198 
199 static int	ptnet_media_change(if_t ifp);
200 static void	ptnet_media_status(if_t ifp, struct ifmediareq *ifmr);
201 #ifdef PTNETMAP_STATS
202 static void	ptnet_tick(void *opaque);
203 #endif
204 
205 static int	ptnet_irqs_init(struct ptnet_softc *sc);
206 static void	ptnet_irqs_fini(struct ptnet_softc *sc);
207 
208 static uint32_t ptnet_nm_ptctl(struct ptnet_softc *sc, uint32_t cmd);
209 static int      ptnet_nm_config(struct netmap_adapter *na,
210 				struct nm_config_info *info);
211 static void	ptnet_update_vnet_hdr(struct ptnet_softc *sc);
212 static int	ptnet_nm_register(struct netmap_adapter *na, int onoff);
213 static int	ptnet_nm_txsync(struct netmap_kring *kring, int flags);
214 static int	ptnet_nm_rxsync(struct netmap_kring *kring, int flags);
215 static void	ptnet_nm_intr(struct netmap_adapter *na, int onoff);
216 
217 static void	ptnet_tx_intr(void *opaque);
218 static void	ptnet_rx_intr(void *opaque);
219 
220 static unsigned	ptnet_rx_discard(struct netmap_kring *kring,
221 				 unsigned int head);
222 static int	ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget,
223 			     bool may_resched);
224 static void	ptnet_rx_task(void *context, int pending);
225 
226 #ifdef DEVICE_POLLING
227 static poll_handler_t ptnet_poll;
228 #endif
229 
230 static device_method_t ptnet_methods[] = {
231 	DEVMETHOD(device_probe,			ptnet_probe),
232 	DEVMETHOD(device_attach,		ptnet_attach),
233 	DEVMETHOD(device_detach,		ptnet_detach),
234 	DEVMETHOD(device_suspend,		ptnet_suspend),
235 	DEVMETHOD(device_resume,		ptnet_resume),
236 	DEVMETHOD(device_shutdown,		ptnet_shutdown),
237 	DEVMETHOD_END
238 };
239 
240 static driver_t ptnet_driver = {
241 	"ptnet",
242 	ptnet_methods,
243 	sizeof(struct ptnet_softc)
244 };
245 
246 /* We use (SI_ORDER_MIDDLE+2) here, see DEV_MODULE_ORDERED() invocation. */
247 DRIVER_MODULE_ORDERED(ptnet, pci, ptnet_driver, NULL, NULL,
248 		      SI_ORDER_MIDDLE + 2);
249 
250 static int
251 ptnet_probe(device_t dev)
252 {
253 	if (pci_get_vendor(dev) != PTNETMAP_PCI_VENDOR_ID ||
254 		pci_get_device(dev) != PTNETMAP_PCI_NETIF_ID) {
255 		return (ENXIO);
256 	}
257 
258 	device_set_desc(dev, "ptnet network adapter");
259 
260 	return (BUS_PROBE_DEFAULT);
261 }
262 
263 static inline void ptnet_kick(struct ptnet_queue *pq)
264 {
265 #ifdef PTNETMAP_STATS
266 	pq->stats.kicks ++;
267 #endif /* PTNETMAP_STATS */
268 	bus_write_4(pq->sc->iomem, pq->kick, 0);
269 }
270 
271 #define PTNET_BUF_RING_SIZE	4096
272 #define PTNET_RX_BUDGET		512
273 #define PTNET_RX_BATCH		1
274 #define PTNET_TX_BUDGET		512
275 #define PTNET_TX_BATCH		64
276 #define PTNET_HDR_SIZE		sizeof(struct virtio_net_hdr_mrg_rxbuf)
277 #define PTNET_MAX_PKT_SIZE	65536
278 
279 #define PTNET_CSUM_OFFLOAD	(CSUM_TCP | CSUM_UDP)
280 #define PTNET_CSUM_OFFLOAD_IPV6	(CSUM_TCP_IPV6 | CSUM_UDP_IPV6)
281 #define PTNET_ALL_OFFLOAD	(CSUM_TSO | PTNET_CSUM_OFFLOAD |\
282 				 PTNET_CSUM_OFFLOAD_IPV6)
283 
284 static int
285 ptnet_attach(device_t dev)
286 {
287 	uint32_t ptfeatures = 0;
288 	unsigned int num_rx_rings, num_tx_rings;
289 	struct netmap_adapter na_arg;
290 	unsigned int nifp_offset;
291 	struct ptnet_softc *sc;
292 	if_t ifp;
293 	uint32_t macreg;
294 	int err, rid;
295 	int i;
296 
297 	sc = device_get_softc(dev);
298 	sc->dev = dev;
299 
300 	/* Setup PCI resources. */
301 	pci_enable_busmaster(dev);
302 
303 	rid = PCIR_BAR(PTNETMAP_IO_PCI_BAR);
304 	sc->iomem = bus_alloc_resource_any(dev, SYS_RES_IOPORT, &rid,
305 					   RF_ACTIVE);
306 	if (sc->iomem == NULL) {
307 		device_printf(dev, "Failed to map I/O BAR\n");
308 		return (ENXIO);
309 	}
310 
311 	/* Negotiate features with the hypervisor. */
312 	if (ptnet_vnet_hdr) {
313 		ptfeatures |= PTNETMAP_F_VNET_HDR;
314 	}
315 	bus_write_4(sc->iomem, PTNET_IO_PTFEAT, ptfeatures); /* wanted */
316 	ptfeatures = bus_read_4(sc->iomem, PTNET_IO_PTFEAT); /* acked */
317 	sc->ptfeatures = ptfeatures;
318 
319 	num_tx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_RINGS);
320 	num_rx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_RINGS);
321 	sc->num_rings = num_tx_rings + num_rx_rings;
322 	sc->num_tx_rings = num_tx_rings;
323 
324 	if (sc->num_rings * sizeof(struct nm_csb_atok) > PAGE_SIZE) {
325 		device_printf(dev, "CSB cannot handle that many rings (%u)\n",
326 				sc->num_rings);
327 		err = ENOMEM;
328 		goto err_path;
329 	}
330 
331 	/* Allocate CSB and carry out CSB allocation protocol. */
332 	sc->csb_gh = contigmalloc(2*PAGE_SIZE, M_DEVBUF, M_NOWAIT | M_ZERO,
333 				  (size_t)0, -1UL, PAGE_SIZE, 0);
334 	if (sc->csb_gh == NULL) {
335 		device_printf(dev, "Failed to allocate CSB\n");
336 		err = ENOMEM;
337 		goto err_path;
338 	}
339 	sc->csb_hg = (struct nm_csb_ktoa *)(((char *)sc->csb_gh) + PAGE_SIZE);
340 
341 	{
342 		/*
343 		 * We use uint64_t rather than vm_paddr_t since we
344 		 * need 64 bit addresses even on 32 bit platforms.
345 		 */
346 		uint64_t paddr = vtophys(sc->csb_gh);
347 
348 		/* CSB allocation protocol: write to BAH first, then
349 		 * to BAL (for both GH and HG sections). */
350 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAH,
351 				(paddr >> 32) & 0xffffffff);
352 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAL,
353 				paddr & 0xffffffff);
354 		paddr = vtophys(sc->csb_hg);
355 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAH,
356 				(paddr >> 32) & 0xffffffff);
357 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAL,
358 				paddr & 0xffffffff);
359 	}
360 
361 	/* Allocate and initialize per-queue data structures. */
362 	sc->queues = malloc(sizeof(struct ptnet_queue) * sc->num_rings,
363 			    M_DEVBUF, M_NOWAIT | M_ZERO);
364 	if (sc->queues == NULL) {
365 		err = ENOMEM;
366 		goto err_path;
367 	}
368 	sc->rxqueues = sc->queues + num_tx_rings;
369 
370 	for (i = 0; i < sc->num_rings; i++) {
371 		struct ptnet_queue *pq = sc->queues + i;
372 
373 		pq->sc = sc;
374 		pq->kring_id = i;
375 		pq->kick = PTNET_IO_KICK_BASE + 4 * i;
376 		pq->atok = sc->csb_gh + i;
377 		pq->ktoa = sc->csb_hg + i;
378 		snprintf(pq->lock_name, sizeof(pq->lock_name), "%s-%d",
379 			 device_get_nameunit(dev), i);
380 		mtx_init(&pq->lock, pq->lock_name, NULL, MTX_DEF);
381 		if (i >= num_tx_rings) {
382 			/* RX queue: fix kring_id. */
383 			pq->kring_id -= num_tx_rings;
384 		} else {
385 			/* TX queue: allocate buf_ring. */
386 			pq->bufring = buf_ring_alloc(PTNET_BUF_RING_SIZE,
387 						M_DEVBUF, M_NOWAIT, &pq->lock);
388 			if (pq->bufring == NULL) {
389 				err = ENOMEM;
390 				goto err_path;
391 			}
392 		}
393 	}
394 
395 	sc->min_tx_space = 64; /* Safe initial value. */
396 
397 	err = ptnet_irqs_init(sc);
398 	if (err) {
399 		goto err_path;
400 	}
401 
402 	/* Setup Ethernet interface. */
403 	sc->ifp = ifp = if_alloc(IFT_ETHER);
404 	if (ifp == NULL) {
405 		device_printf(dev, "Failed to allocate ifnet\n");
406 		err = ENOMEM;
407 		goto err_path;
408 	}
409 
410 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
411 	ifp->if_baudrate = IF_Gbps(10);
412 	ifp->if_softc = sc;
413 	ifp->if_flags = IFF_BROADCAST | IFF_MULTICAST | IFF_SIMPLEX;
414 	ifp->if_init = ptnet_init;
415 	ifp->if_ioctl = ptnet_ioctl;
416 	ifp->if_get_counter = ptnet_get_counter;
417 	ifp->if_transmit = ptnet_transmit;
418 	ifp->if_qflush = ptnet_qflush;
419 
420 	ifmedia_init(&sc->media, IFM_IMASK, ptnet_media_change,
421 		     ptnet_media_status);
422 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T | IFM_FDX, 0, NULL);
423 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T | IFM_FDX);
424 
425 	macreg = bus_read_4(sc->iomem, PTNET_IO_MAC_HI);
426 	sc->hwaddr[0] = (macreg >> 8) & 0xff;
427 	sc->hwaddr[1] = macreg & 0xff;
428 	macreg = bus_read_4(sc->iomem, PTNET_IO_MAC_LO);
429 	sc->hwaddr[2] = (macreg >> 24) & 0xff;
430 	sc->hwaddr[3] = (macreg >> 16) & 0xff;
431 	sc->hwaddr[4] = (macreg >> 8) & 0xff;
432 	sc->hwaddr[5] = macreg & 0xff;
433 
434 	ether_ifattach(ifp, sc->hwaddr);
435 
436 	ifp->if_hdrlen = sizeof(struct ether_vlan_header);
437 	ifp->if_capabilities |= IFCAP_JUMBO_MTU | IFCAP_VLAN_MTU;
438 
439 	if (sc->ptfeatures & PTNETMAP_F_VNET_HDR) {
440 		/* Similarly to what the vtnet driver does, we can emulate
441 		 * VLAN offloadings by inserting and removing the 802.1Q
442 		 * header during transmit and receive. We are then able
443 		 * to do checksum offloading of VLAN frames. */
444 		ifp->if_capabilities |= IFCAP_HWCSUM | IFCAP_HWCSUM_IPV6
445 					| IFCAP_VLAN_HWCSUM
446 					| IFCAP_TSO | IFCAP_LRO
447 					| IFCAP_VLAN_HWTSO
448 					| IFCAP_VLAN_HWTAGGING;
449 	}
450 
451 	ifp->if_capenable = ifp->if_capabilities;
452 #ifdef DEVICE_POLLING
453 	/* Don't enable polling by default. */
454 	ifp->if_capabilities |= IFCAP_POLLING;
455 #endif
456 	snprintf(sc->lock_name, sizeof(sc->lock_name),
457 		 "%s", device_get_nameunit(dev));
458 	mtx_init(&sc->lock, sc->lock_name, "ptnet core lock", MTX_DEF);
459 	callout_init_mtx(&sc->tick, &sc->lock, 0);
460 
461 	/* Prepare a netmap_adapter struct instance to do netmap_attach(). */
462 	nifp_offset = bus_read_4(sc->iomem, PTNET_IO_NIFP_OFS);
463 	memset(&na_arg, 0, sizeof(na_arg));
464 	na_arg.ifp = ifp;
465 	na_arg.num_tx_desc = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_SLOTS);
466 	na_arg.num_rx_desc = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_SLOTS);
467 	na_arg.num_tx_rings = num_tx_rings;
468 	na_arg.num_rx_rings = num_rx_rings;
469 	na_arg.nm_config = ptnet_nm_config;
470 	na_arg.nm_krings_create = ptnet_nm_krings_create;
471 	na_arg.nm_krings_delete = ptnet_nm_krings_delete;
472 	na_arg.nm_dtor = ptnet_nm_dtor;
473 	na_arg.nm_intr = ptnet_nm_intr;
474 	na_arg.nm_register = ptnet_nm_register;
475 	na_arg.nm_txsync = ptnet_nm_txsync;
476 	na_arg.nm_rxsync = ptnet_nm_rxsync;
477 
478 	netmap_pt_guest_attach(&na_arg, nifp_offset,
479                                 bus_read_4(sc->iomem, PTNET_IO_HOSTMEMID));
480 
481 	/* Now a netmap adapter for this ifp has been allocated, and it
482 	 * can be accessed through NA(ifp). We also have to initialize the CSB
483 	 * pointer. */
484 	sc->ptna = (struct netmap_pt_guest_adapter *)NA(ifp);
485 
486 	/* If virtio-net header was negotiated, set the virt_hdr_len field in
487 	 * the netmap adapter, to inform users that this netmap adapter requires
488 	 * the application to deal with the headers. */
489 	ptnet_update_vnet_hdr(sc);
490 
491 	device_printf(dev, "%s() completed\n", __func__);
492 
493 	return (0);
494 
495 err_path:
496 	ptnet_detach(dev);
497 	return err;
498 }
499 
500 /* Stop host sync-kloop if it was running. */
501 static void
502 ptnet_device_shutdown(struct ptnet_softc *sc)
503 {
504 	ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_DELETE);
505 	bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAH, 0);
506 	bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAL, 0);
507 	bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAH, 0);
508 	bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAL, 0);
509 }
510 
511 static int
512 ptnet_detach(device_t dev)
513 {
514 	struct ptnet_softc *sc = device_get_softc(dev);
515 	int i;
516 
517 	ptnet_device_shutdown(sc);
518 
519 #ifdef DEVICE_POLLING
520 	if (sc->ifp->if_capenable & IFCAP_POLLING) {
521 		ether_poll_deregister(sc->ifp);
522 	}
523 #endif
524 	callout_drain(&sc->tick);
525 
526 	if (sc->queues) {
527 		/* Drain taskqueues before calling if_detach. */
528 		for (i = 0; i < sc->num_rings; i++) {
529 			struct ptnet_queue *pq = sc->queues + i;
530 
531 			if (pq->taskq) {
532 				taskqueue_drain(pq->taskq, &pq->task);
533 			}
534 		}
535 	}
536 
537 	if (sc->ifp) {
538 		ether_ifdetach(sc->ifp);
539 
540 		/* Uninitialize netmap adapters for this device. */
541 		netmap_detach(sc->ifp);
542 
543 		ifmedia_removeall(&sc->media);
544 		if_free(sc->ifp);
545 		sc->ifp = NULL;
546 	}
547 
548 	ptnet_irqs_fini(sc);
549 
550 	if (sc->csb_gh) {
551 		contigfree(sc->csb_gh, 2*PAGE_SIZE, M_DEVBUF);
552 		sc->csb_gh = NULL;
553 		sc->csb_hg = NULL;
554 	}
555 
556 	if (sc->queues) {
557 		for (i = 0; i < sc->num_rings; i++) {
558 			struct ptnet_queue *pq = sc->queues + i;
559 
560 			if (mtx_initialized(&pq->lock)) {
561 				mtx_destroy(&pq->lock);
562 			}
563 			if (pq->bufring != NULL) {
564 				buf_ring_free(pq->bufring, M_DEVBUF);
565 			}
566 		}
567 		free(sc->queues, M_DEVBUF);
568 		sc->queues = NULL;
569 	}
570 
571 	if (sc->iomem) {
572 		bus_release_resource(dev, SYS_RES_IOPORT,
573 				     PCIR_BAR(PTNETMAP_IO_PCI_BAR), sc->iomem);
574 		sc->iomem = NULL;
575 	}
576 
577 	mtx_destroy(&sc->lock);
578 
579 	device_printf(dev, "%s() completed\n", __func__);
580 
581 	return (0);
582 }
583 
584 static int
585 ptnet_suspend(device_t dev)
586 {
587 	struct ptnet_softc *sc = device_get_softc(dev);
588 
589 	(void)sc;
590 
591 	return (0);
592 }
593 
594 static int
595 ptnet_resume(device_t dev)
596 {
597 	struct ptnet_softc *sc = device_get_softc(dev);
598 
599 	(void)sc;
600 
601 	return (0);
602 }
603 
604 static int
605 ptnet_shutdown(device_t dev)
606 {
607 	struct ptnet_softc *sc = device_get_softc(dev);
608 
609 	ptnet_device_shutdown(sc);
610 
611 	return (0);
612 }
613 
614 static int
615 ptnet_irqs_init(struct ptnet_softc *sc)
616 {
617 	int rid = PCIR_BAR(PTNETMAP_MSIX_PCI_BAR);
618 	int nvecs = sc->num_rings;
619 	device_t dev = sc->dev;
620 	int err = ENOSPC;
621 	int cpu_cur;
622 	int i;
623 
624 	if (pci_find_cap(dev, PCIY_MSIX, NULL) != 0)  {
625 		device_printf(dev, "Could not find MSI-X capability\n");
626 		return (ENXIO);
627 	}
628 
629 	sc->msix_mem = bus_alloc_resource_any(dev, SYS_RES_MEMORY,
630 					      &rid, RF_ACTIVE);
631 	if (sc->msix_mem == NULL) {
632 		device_printf(dev, "Failed to allocate MSIX PCI BAR\n");
633 		return (ENXIO);
634 	}
635 
636 	if (pci_msix_count(dev) < nvecs) {
637 		device_printf(dev, "Not enough MSI-X vectors\n");
638 		goto err_path;
639 	}
640 
641 	err = pci_alloc_msix(dev, &nvecs);
642 	if (err) {
643 		device_printf(dev, "Failed to allocate MSI-X vectors\n");
644 		goto err_path;
645 	}
646 
647 	for (i = 0; i < nvecs; i++) {
648 		struct ptnet_queue *pq = sc->queues + i;
649 
650 		rid = i + 1;
651 		pq->irq = bus_alloc_resource_any(dev, SYS_RES_IRQ, &rid,
652 						 RF_ACTIVE);
653 		if (pq->irq == NULL) {
654 			device_printf(dev, "Failed to allocate interrupt "
655 					   "for queue #%d\n", i);
656 			err = ENOSPC;
657 			goto err_path;
658 		}
659 	}
660 
661 	cpu_cur = CPU_FIRST();
662 	for (i = 0; i < nvecs; i++) {
663 		struct ptnet_queue *pq = sc->queues + i;
664 		void (*handler)(void *) = ptnet_tx_intr;
665 
666 		if (i >= sc->num_tx_rings) {
667 			handler = ptnet_rx_intr;
668 		}
669 		err = bus_setup_intr(dev, pq->irq, INTR_TYPE_NET | INTR_MPSAFE,
670 				     NULL /* intr_filter */, handler,
671 				     pq, &pq->cookie);
672 		if (err) {
673 			device_printf(dev, "Failed to register intr handler "
674 					   "for queue #%d\n", i);
675 			goto err_path;
676 		}
677 
678 		bus_describe_intr(dev, pq->irq, pq->cookie, "q%d", i);
679 #if 0
680 		bus_bind_intr(sc->dev, pq->irq, cpu_cur);
681 #endif
682 		cpu_cur = CPU_NEXT(cpu_cur);
683 	}
684 
685 	device_printf(dev, "Allocated %d MSI-X vectors\n", nvecs);
686 
687 	cpu_cur = CPU_FIRST();
688 	for (i = 0; i < nvecs; i++) {
689 		struct ptnet_queue *pq = sc->queues + i;
690 
691 		if (i < sc->num_tx_rings)
692 			TASK_INIT(&pq->task, 0, ptnet_tx_task, pq);
693 		else
694 			NET_TASK_INIT(&pq->task, 0, ptnet_rx_task, pq);
695 
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->atok->appl_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->atok->appl_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_users == 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_users == 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 static uint64_t
1011 ptnet_get_counter(if_t ifp, ift_counter cnt)
1012 {
1013 	struct ptnet_softc *sc = if_getsoftc(ifp);
1014 	struct ptnet_queue_stats stats[2];
1015 	int i;
1016 
1017 	/* Accumulate statistics over the queues. */
1018 	memset(stats, 0, sizeof(stats));
1019 	for (i = 0; i < sc->num_rings; i++) {
1020 		struct ptnet_queue *pq = sc->queues + i;
1021 		int idx = (i < sc->num_tx_rings) ? 0 : 1;
1022 
1023 		stats[idx].packets	+= pq->stats.packets;
1024 		stats[idx].bytes	+= pq->stats.bytes;
1025 		stats[idx].errors	+= pq->stats.errors;
1026 		stats[idx].iqdrops	+= pq->stats.iqdrops;
1027 		stats[idx].mcasts	+= pq->stats.mcasts;
1028 	}
1029 
1030 	switch (cnt) {
1031 	case IFCOUNTER_IPACKETS:
1032 		return (stats[1].packets);
1033 	case IFCOUNTER_IQDROPS:
1034 		return (stats[1].iqdrops);
1035 	case IFCOUNTER_IERRORS:
1036 		return (stats[1].errors);
1037 	case IFCOUNTER_OPACKETS:
1038 		return (stats[0].packets);
1039 	case IFCOUNTER_OBYTES:
1040 		return (stats[0].bytes);
1041 	case IFCOUNTER_OMCASTS:
1042 		return (stats[0].mcasts);
1043 	default:
1044 		return (if_get_counter_default(ifp, cnt));
1045 	}
1046 }
1047 
1048 
1049 #ifdef PTNETMAP_STATS
1050 /* Called under core lock. */
1051 static void
1052 ptnet_tick(void *opaque)
1053 {
1054 	struct ptnet_softc *sc = opaque;
1055 	int i;
1056 
1057 	for (i = 0; i < sc->num_rings; i++) {
1058 		struct ptnet_queue *pq = sc->queues + i;
1059 		struct ptnet_queue_stats cur = pq->stats;
1060 		struct timeval now;
1061 		unsigned int delta;
1062 
1063 		microtime(&now);
1064 		delta = now.tv_usec - sc->last_ts.tv_usec +
1065 			(now.tv_sec - sc->last_ts.tv_sec) * 1000000;
1066 		delta /= 1000; /* in milliseconds */
1067 
1068 		if (delta == 0)
1069 			continue;
1070 
1071 		device_printf(sc->dev, "#%d[%u ms]:pkts %lu, kicks %lu, "
1072 			      "intr %lu\n", i, delta,
1073 			      (cur.packets - pq->last_stats.packets),
1074 			      (cur.kicks - pq->last_stats.kicks),
1075 			      (cur.intrs - pq->last_stats.intrs));
1076 		pq->last_stats = cur;
1077 	}
1078 	microtime(&sc->last_ts);
1079 	callout_schedule(&sc->tick, hz);
1080 }
1081 #endif /* PTNETMAP_STATS */
1082 
1083 static void
1084 ptnet_media_status(if_t ifp, struct ifmediareq *ifmr)
1085 {
1086 	/* We are always active, as the backend netmap port is
1087 	 * always open in netmap mode. */
1088 	ifmr->ifm_status = IFM_AVALID | IFM_ACTIVE;
1089 	ifmr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
1090 }
1091 
1092 static uint32_t
1093 ptnet_nm_ptctl(struct ptnet_softc *sc, uint32_t cmd)
1094 {
1095 	/*
1096 	 * Write a command and read back error status,
1097 	 * with zero meaning success.
1098 	 */
1099 	bus_write_4(sc->iomem, PTNET_IO_PTCTL, cmd);
1100 	return bus_read_4(sc->iomem, PTNET_IO_PTCTL);
1101 }
1102 
1103 static int
1104 ptnet_nm_config(struct netmap_adapter *na, struct nm_config_info *info)
1105 {
1106 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1107 
1108 	info->num_tx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_RINGS);
1109 	info->num_rx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_RINGS);
1110 	info->num_tx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_SLOTS);
1111 	info->num_rx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_SLOTS);
1112 	info->rx_buf_maxsize = NETMAP_BUF_SIZE(na);
1113 
1114 	device_printf(sc->dev, "txr %u, rxr %u, txd %u, rxd %u, rxbufsz %u\n",
1115 			info->num_tx_rings, info->num_rx_rings,
1116 			info->num_tx_descs, info->num_rx_descs,
1117 			info->rx_buf_maxsize);
1118 
1119 	return 0;
1120 }
1121 
1122 static void
1123 ptnet_sync_from_csb(struct ptnet_softc *sc, struct netmap_adapter *na)
1124 {
1125 	int i;
1126 
1127 	/* Sync krings from the host, reading from
1128 	 * CSB. */
1129 	for (i = 0; i < sc->num_rings; i++) {
1130 		struct nm_csb_atok *atok = sc->queues[i].atok;
1131 		struct nm_csb_ktoa *ktoa = sc->queues[i].ktoa;
1132 		struct netmap_kring *kring;
1133 
1134 		if (i < na->num_tx_rings) {
1135 			kring = na->tx_rings[i];
1136 		} else {
1137 			kring = na->rx_rings[i - na->num_tx_rings];
1138 		}
1139 		kring->rhead = kring->ring->head = atok->head;
1140 		kring->rcur = kring->ring->cur = atok->cur;
1141 		kring->nr_hwcur = ktoa->hwcur;
1142 		kring->nr_hwtail = kring->rtail =
1143 			kring->ring->tail = ktoa->hwtail;
1144 
1145 		nm_prdis("%d,%d: csb {hc %u h %u c %u ht %u}", t, i,
1146 		   ktoa->hwcur, atok->head, atok->cur,
1147 		   ktoa->hwtail);
1148 		nm_prdis("%d,%d: kring {hc %u rh %u rc %u h %u c %u ht %u rt %u t %u}",
1149 		   t, i, kring->nr_hwcur, kring->rhead, kring->rcur,
1150 		   kring->ring->head, kring->ring->cur, kring->nr_hwtail,
1151 		   kring->rtail, kring->ring->tail);
1152 	}
1153 }
1154 
1155 static void
1156 ptnet_update_vnet_hdr(struct ptnet_softc *sc)
1157 {
1158 	unsigned int wanted_hdr_len = ptnet_vnet_hdr ? PTNET_HDR_SIZE : 0;
1159 
1160 	bus_write_4(sc->iomem, PTNET_IO_VNET_HDR_LEN, wanted_hdr_len);
1161 	sc->vnet_hdr_len = bus_read_4(sc->iomem, PTNET_IO_VNET_HDR_LEN);
1162 	sc->ptna->hwup.up.virt_hdr_len = sc->vnet_hdr_len;
1163 }
1164 
1165 static int
1166 ptnet_nm_register(struct netmap_adapter *na, int onoff)
1167 {
1168 	/* device-specific */
1169 	if_t ifp = na->ifp;
1170 	struct ptnet_softc *sc = if_getsoftc(ifp);
1171 	int native = (na == &sc->ptna->hwup.up);
1172 	struct ptnet_queue *pq;
1173 	int ret = 0;
1174 	int i;
1175 
1176 	if (!onoff) {
1177 		sc->ptna->backend_users--;
1178 	}
1179 
1180 	/* If this is the last netmap client, guest interrupt enable flags may
1181 	 * be in arbitrary state. Since these flags are going to be used also
1182 	 * by the netdevice driver, we have to make sure to start with
1183 	 * notifications enabled. Also, schedule NAPI to flush pending packets
1184 	 * in the RX rings, since we will not receive further interrupts
1185 	 * until these will be processed. */
1186 	if (native && !onoff && na->active_fds == 0) {
1187 		nm_prinf("Exit netmap mode, re-enable interrupts");
1188 		for (i = 0; i < sc->num_rings; i++) {
1189 			pq = sc->queues + i;
1190 			pq->atok->appl_need_kick = 1;
1191 		}
1192 	}
1193 
1194 	if (onoff) {
1195 		if (sc->ptna->backend_users == 0) {
1196 			/* Initialize notification enable fields in the CSB. */
1197 			for (i = 0; i < sc->num_rings; i++) {
1198 				pq = sc->queues + i;
1199 				pq->ktoa->kern_need_kick = 1;
1200 				pq->atok->appl_need_kick =
1201 					(!(ifp->if_capenable & IFCAP_POLLING)
1202 						&& i >= sc->num_tx_rings);
1203 			}
1204 
1205 			/* Set the virtio-net header length. */
1206 			ptnet_update_vnet_hdr(sc);
1207 
1208 			/* Make sure the host adapter passed through is ready
1209 			 * for txsync/rxsync. */
1210 			ret = ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_CREATE);
1211 			if (ret) {
1212 				return ret;
1213 			}
1214 
1215 			/* Align the guest krings and rings to the state stored
1216 			 * in the CSB. */
1217 			ptnet_sync_from_csb(sc, na);
1218 		}
1219 
1220 		/* If not native, don't call nm_set_native_flags, since we don't want
1221 		 * to replace if_transmit method, nor set NAF_NETMAP_ON */
1222 		if (native) {
1223 			netmap_krings_mode_commit(na, onoff);
1224 			nm_set_native_flags(na);
1225 		}
1226 
1227 	} else {
1228 		if (native) {
1229 			nm_clear_native_flags(na);
1230 			netmap_krings_mode_commit(na, onoff);
1231 		}
1232 
1233 		if (sc->ptna->backend_users == 0) {
1234 			ret = ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_DELETE);
1235 		}
1236 	}
1237 
1238 	if (onoff) {
1239 		sc->ptna->backend_users++;
1240 	}
1241 
1242 	return ret;
1243 }
1244 
1245 static int
1246 ptnet_nm_txsync(struct netmap_kring *kring, int flags)
1247 {
1248 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1249 	struct ptnet_queue *pq = sc->queues + kring->ring_id;
1250 	bool notify;
1251 
1252 	notify = netmap_pt_guest_txsync(pq->atok, pq->ktoa, kring, flags);
1253 	if (notify) {
1254 		ptnet_kick(pq);
1255 	}
1256 
1257 	return 0;
1258 }
1259 
1260 static int
1261 ptnet_nm_rxsync(struct netmap_kring *kring, int flags)
1262 {
1263 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1264 	struct ptnet_queue *pq = sc->rxqueues + kring->ring_id;
1265 	bool notify;
1266 
1267 	notify = netmap_pt_guest_rxsync(pq->atok, pq->ktoa, kring, flags);
1268 	if (notify) {
1269 		ptnet_kick(pq);
1270 	}
1271 
1272 	return 0;
1273 }
1274 
1275 static void
1276 ptnet_nm_intr(struct netmap_adapter *na, int onoff)
1277 {
1278 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1279 	int i;
1280 
1281 	for (i = 0; i < sc->num_rings; i++) {
1282 		struct ptnet_queue *pq = sc->queues + i;
1283 		pq->atok->appl_need_kick = onoff;
1284 	}
1285 }
1286 
1287 static void
1288 ptnet_tx_intr(void *opaque)
1289 {
1290 	struct ptnet_queue *pq = opaque;
1291 	struct ptnet_softc *sc = pq->sc;
1292 
1293 	DBG(device_printf(sc->dev, "Tx interrupt #%d\n", pq->kring_id));
1294 #ifdef PTNETMAP_STATS
1295 	pq->stats.intrs ++;
1296 #endif /* PTNETMAP_STATS */
1297 
1298 	if (netmap_tx_irq(sc->ifp, pq->kring_id) != NM_IRQ_PASS) {
1299 		return;
1300 	}
1301 
1302 	/* Schedule the tasqueue to flush process transmissions requests.
1303 	 * However, vtnet, if_em and if_igb just call ptnet_transmit() here,
1304 	 * at least when using MSI-X interrupts. The if_em driver, instead
1305 	 * schedule taskqueue when using legacy interrupts. */
1306 	taskqueue_enqueue(pq->taskq, &pq->task);
1307 }
1308 
1309 static void
1310 ptnet_rx_intr(void *opaque)
1311 {
1312 	struct ptnet_queue *pq = opaque;
1313 	struct ptnet_softc *sc = pq->sc;
1314 	unsigned int unused;
1315 
1316 	DBG(device_printf(sc->dev, "Rx interrupt #%d\n", pq->kring_id));
1317 #ifdef PTNETMAP_STATS
1318 	pq->stats.intrs ++;
1319 #endif /* PTNETMAP_STATS */
1320 
1321 	if (netmap_rx_irq(sc->ifp, pq->kring_id, &unused) != NM_IRQ_PASS) {
1322 		return;
1323 	}
1324 
1325 	/* Like vtnet, if_igb and if_em drivers when using MSI-X interrupts,
1326 	 * receive-side processing is executed directly in the interrupt
1327 	 * service routine. Alternatively, we may schedule the taskqueue. */
1328 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
1329 }
1330 
1331 static void
1332 ptnet_vlan_tag_remove(struct mbuf *m)
1333 {
1334 	struct ether_vlan_header *evh;
1335 
1336 	evh = mtod(m, struct ether_vlan_header *);
1337 	m->m_pkthdr.ether_vtag = ntohs(evh->evl_tag);
1338 	m->m_flags |= M_VLANTAG;
1339 
1340 	/* Strip the 802.1Q header. */
1341 	bcopy((char *) evh, (char *) evh + ETHER_VLAN_ENCAP_LEN,
1342 	    ETHER_HDR_LEN - ETHER_TYPE_LEN);
1343 	m_adj(m, ETHER_VLAN_ENCAP_LEN);
1344 }
1345 
1346 static void
1347 ptnet_ring_update(struct ptnet_queue *pq, struct netmap_kring *kring,
1348 		  unsigned int head, unsigned int sync_flags)
1349 {
1350 	struct netmap_ring *ring = kring->ring;
1351 	struct nm_csb_atok *atok = pq->atok;
1352 	struct nm_csb_ktoa *ktoa = pq->ktoa;
1353 
1354 	/* Some packets have been pushed to the netmap ring. We have
1355 	 * to tell the host to process the new packets, updating cur
1356 	 * and head in the CSB. */
1357 	ring->head = ring->cur = head;
1358 
1359 	/* Mimic nm_txsync_prologue/nm_rxsync_prologue. */
1360 	kring->rcur = kring->rhead = head;
1361 
1362 	nm_sync_kloop_appl_write(atok, kring->rcur, kring->rhead);
1363 
1364 	/* Kick the host if needed. */
1365 	if (NM_ACCESS_ONCE(ktoa->kern_need_kick)) {
1366 		atok->sync_flags = sync_flags;
1367 		ptnet_kick(pq);
1368 	}
1369 }
1370 
1371 #define PTNET_TX_NOSPACE(_h, _k, _min)	\
1372 	((((_h) < (_k)->rtail) ? 0 : (_k)->nkr_num_slots) + \
1373 		(_k)->rtail - (_h)) < (_min)
1374 
1375 /* This function may be called by the network stack, or by
1376  * by the taskqueue thread. */
1377 static int
1378 ptnet_drain_transmit_queue(struct ptnet_queue *pq, unsigned int budget,
1379 			   bool may_resched)
1380 {
1381 	struct ptnet_softc *sc = pq->sc;
1382 	bool have_vnet_hdr = sc->vnet_hdr_len;
1383 	struct netmap_adapter *na = &sc->ptna->dr.up;
1384 	if_t ifp = sc->ifp;
1385 	unsigned int batch_count = 0;
1386 	struct nm_csb_atok *atok;
1387 	struct nm_csb_ktoa *ktoa;
1388 	struct netmap_kring *kring;
1389 	struct netmap_ring *ring;
1390 	struct netmap_slot *slot;
1391 	unsigned int count = 0;
1392 	unsigned int minspace;
1393 	unsigned int head;
1394 	unsigned int lim;
1395 	struct mbuf *mhead;
1396 	struct mbuf *mf;
1397 	int nmbuf_bytes;
1398 	uint8_t *nmbuf;
1399 
1400 	if (!PTNET_Q_TRYLOCK(pq)) {
1401 		/* We failed to acquire the lock, schedule the taskqueue. */
1402 		nm_prlim(1, "Deferring TX work");
1403 		if (may_resched) {
1404 			taskqueue_enqueue(pq->taskq, &pq->task);
1405 		}
1406 
1407 		return 0;
1408 	}
1409 
1410 	if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1411 		PTNET_Q_UNLOCK(pq);
1412 		nm_prlim(1, "Interface is down");
1413 		return ENETDOWN;
1414 	}
1415 
1416 	atok = pq->atok;
1417 	ktoa = pq->ktoa;
1418 	kring = na->tx_rings[pq->kring_id];
1419 	ring = kring->ring;
1420 	lim = kring->nkr_num_slots - 1;
1421 	head = ring->head;
1422 	minspace = sc->min_tx_space;
1423 
1424 	while (count < budget) {
1425 		if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1426 			/* We ran out of slot, let's see if the host has
1427 			 * freed up some, by reading hwcur and hwtail from
1428 			 * the CSB. */
1429 			ptnet_sync_tail(ktoa, kring);
1430 
1431 			if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1432 				/* Still no slots available. Reactivate the
1433 				 * interrupts so that we can be notified
1434 				 * when some free slots are made available by
1435 				 * the host. */
1436 				atok->appl_need_kick = 1;
1437 
1438 				/* Double check. We need a full barrier to
1439 				 * prevent the store to atok->appl_need_kick
1440 				 * to be reordered with the load from
1441 				 * ktoa->hwcur and ktoa->hwtail (store-load
1442 				 * barrier). */
1443 				nm_stld_barrier();
1444 				ptnet_sync_tail(ktoa, kring);
1445 				if (likely(PTNET_TX_NOSPACE(head, kring,
1446 							    minspace))) {
1447 					break;
1448 				}
1449 
1450 				nm_prlim(1, "Found more slots by doublecheck");
1451 				/* More slots were freed before reactivating
1452 				 * the interrupts. */
1453 				atok->appl_need_kick = 0;
1454 			}
1455 		}
1456 
1457 		mhead = drbr_peek(ifp, pq->bufring);
1458 		if (!mhead) {
1459 			break;
1460 		}
1461 
1462 		/* Initialize transmission state variables. */
1463 		slot = ring->slot + head;
1464 		nmbuf = NMB(na, slot);
1465 		nmbuf_bytes = 0;
1466 
1467 		/* If needed, prepare the virtio-net header at the beginning
1468 		 * of the first slot. */
1469 		if (have_vnet_hdr) {
1470 			struct virtio_net_hdr *vh =
1471 					(struct virtio_net_hdr *)nmbuf;
1472 
1473 			/* For performance, we could replace this memset() with
1474 			 * two 8-bytes-wide writes. */
1475 			memset(nmbuf, 0, PTNET_HDR_SIZE);
1476 			if (mhead->m_pkthdr.csum_flags & PTNET_ALL_OFFLOAD) {
1477 				mhead = virtio_net_tx_offload(ifp, mhead, false,
1478 							 vh);
1479 				if (unlikely(!mhead)) {
1480 					/* Packet dropped because errors
1481 					 * occurred while preparing the vnet
1482 					 * header. Let's go ahead with the next
1483 					 * packet. */
1484 					pq->stats.errors ++;
1485 					drbr_advance(ifp, pq->bufring);
1486 					continue;
1487 				}
1488 			}
1489 			nm_prdis(1, "%s: [csum_flags %lX] vnet hdr: flags %x "
1490 			      "csum_start %u csum_ofs %u hdr_len = %u "
1491 			      "gso_size %u gso_type %x", __func__,
1492 			      mhead->m_pkthdr.csum_flags, vh->flags,
1493 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
1494 			      vh->gso_size, vh->gso_type);
1495 
1496 			nmbuf += PTNET_HDR_SIZE;
1497 			nmbuf_bytes += PTNET_HDR_SIZE;
1498 		}
1499 
1500 		for (mf = mhead; mf; mf = mf->m_next) {
1501 			uint8_t *mdata = mf->m_data;
1502 			int mlen = mf->m_len;
1503 
1504 			for (;;) {
1505 				int copy = NETMAP_BUF_SIZE(na) - nmbuf_bytes;
1506 
1507 				if (mlen < copy) {
1508 					copy = mlen;
1509 				}
1510 				memcpy(nmbuf, mdata, copy);
1511 
1512 				mdata += copy;
1513 				mlen -= copy;
1514 				nmbuf += copy;
1515 				nmbuf_bytes += copy;
1516 
1517 				if (!mlen) {
1518 					break;
1519 				}
1520 
1521 				slot->len = nmbuf_bytes;
1522 				slot->flags = NS_MOREFRAG;
1523 
1524 				head = nm_next(head, lim);
1525 				KASSERT(head != ring->tail,
1526 					("Unexpectedly run out of TX space"));
1527 				slot = ring->slot + head;
1528 				nmbuf = NMB(na, slot);
1529 				nmbuf_bytes = 0;
1530 			}
1531 		}
1532 
1533 		/* Complete last slot and update head. */
1534 		slot->len = nmbuf_bytes;
1535 		slot->flags = 0;
1536 		head = nm_next(head, lim);
1537 
1538 		/* Consume the packet just processed. */
1539 		drbr_advance(ifp, pq->bufring);
1540 
1541 		/* Copy the packet to listeners. */
1542 		ETHER_BPF_MTAP(ifp, mhead);
1543 
1544 		pq->stats.packets ++;
1545 		pq->stats.bytes += mhead->m_pkthdr.len;
1546 		if (mhead->m_flags & M_MCAST) {
1547 			pq->stats.mcasts ++;
1548 		}
1549 
1550 		m_freem(mhead);
1551 
1552 		count ++;
1553 		if (++batch_count == PTNET_TX_BATCH) {
1554 			ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1555 			batch_count = 0;
1556 		}
1557 	}
1558 
1559 	if (batch_count) {
1560 		ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1561 	}
1562 
1563 	if (count >= budget && may_resched) {
1564 		DBG(nm_prlim(1, "out of budget: resched, %d mbufs pending\n",
1565 					drbr_inuse(ifp, pq->bufring)));
1566 		taskqueue_enqueue(pq->taskq, &pq->task);
1567 	}
1568 
1569 	PTNET_Q_UNLOCK(pq);
1570 
1571 	return count;
1572 }
1573 
1574 static int
1575 ptnet_transmit(if_t ifp, struct mbuf *m)
1576 {
1577 	struct ptnet_softc *sc = if_getsoftc(ifp);
1578 	struct ptnet_queue *pq;
1579 	unsigned int queue_idx;
1580 	int err;
1581 
1582 	DBG(device_printf(sc->dev, "transmit %p\n", m));
1583 
1584 	/* Insert 802.1Q header if needed. */
1585 	if (m->m_flags & M_VLANTAG) {
1586 		m = ether_vlanencap(m, m->m_pkthdr.ether_vtag);
1587 		if (m == NULL) {
1588 			return ENOBUFS;
1589 		}
1590 		m->m_flags &= ~M_VLANTAG;
1591 	}
1592 
1593 	/* Get the flow-id if available. */
1594 	queue_idx = (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) ?
1595 		    m->m_pkthdr.flowid : curcpu;
1596 
1597 	if (unlikely(queue_idx >= sc->num_tx_rings)) {
1598 		queue_idx %= sc->num_tx_rings;
1599 	}
1600 
1601 	pq = sc->queues + queue_idx;
1602 
1603 	err = drbr_enqueue(ifp, pq->bufring, m);
1604 	if (err) {
1605 		/* ENOBUFS when the bufring is full */
1606 		nm_prlim(1, "%s: drbr_enqueue() failed %d\n",
1607 			__func__, err);
1608 		pq->stats.errors ++;
1609 		return err;
1610 	}
1611 
1612 	if (ifp->if_capenable & IFCAP_POLLING) {
1613 		/* If polling is on, the transmit queues will be
1614 		 * drained by the poller. */
1615 		return 0;
1616 	}
1617 
1618 	err = ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
1619 
1620 	return (err < 0) ? err : 0;
1621 }
1622 
1623 static unsigned int
1624 ptnet_rx_discard(struct netmap_kring *kring, unsigned int head)
1625 {
1626 	struct netmap_ring *ring = kring->ring;
1627 	struct netmap_slot *slot = ring->slot + head;
1628 
1629 	for (;;) {
1630 		head = nm_next(head, kring->nkr_num_slots - 1);
1631 		if (!(slot->flags & NS_MOREFRAG) || head == ring->tail) {
1632 			break;
1633 		}
1634 		slot = ring->slot + head;
1635 	}
1636 
1637 	return head;
1638 }
1639 
1640 static inline struct mbuf *
1641 ptnet_rx_slot(struct mbuf *mtail, uint8_t *nmbuf, unsigned int nmbuf_len)
1642 {
1643 	uint8_t *mdata = mtod(mtail, uint8_t *) + mtail->m_len;
1644 
1645 	do {
1646 		unsigned int copy;
1647 
1648 		if (mtail->m_len == MCLBYTES) {
1649 			struct mbuf *mf;
1650 
1651 			mf = m_getcl(M_NOWAIT, MT_DATA, 0);
1652 			if (unlikely(!mf)) {
1653 				return NULL;
1654 			}
1655 
1656 			mtail->m_next = mf;
1657 			mtail = mf;
1658 			mdata = mtod(mtail, uint8_t *);
1659 			mtail->m_len = 0;
1660 		}
1661 
1662 		copy = MCLBYTES - mtail->m_len;
1663 		if (nmbuf_len < copy) {
1664 			copy = nmbuf_len;
1665 		}
1666 
1667 		memcpy(mdata, nmbuf, copy);
1668 
1669 		nmbuf += copy;
1670 		nmbuf_len -= copy;
1671 		mdata += copy;
1672 		mtail->m_len += copy;
1673 	} while (nmbuf_len);
1674 
1675 	return mtail;
1676 }
1677 
1678 static int
1679 ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget, bool may_resched)
1680 {
1681 	struct ptnet_softc *sc = pq->sc;
1682 	bool have_vnet_hdr = sc->vnet_hdr_len;
1683 	struct nm_csb_atok *atok = pq->atok;
1684 	struct nm_csb_ktoa *ktoa = pq->ktoa;
1685 	struct netmap_adapter *na = &sc->ptna->dr.up;
1686 	struct netmap_kring *kring = na->rx_rings[pq->kring_id];
1687 	struct netmap_ring *ring = kring->ring;
1688 	unsigned int const lim = kring->nkr_num_slots - 1;
1689 	unsigned int batch_count = 0;
1690 	if_t ifp = sc->ifp;
1691 	unsigned int count = 0;
1692 	uint32_t head;
1693 
1694 	PTNET_Q_LOCK(pq);
1695 
1696 	if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1697 		goto unlock;
1698 	}
1699 
1700 	kring->nr_kflags &= ~NKR_PENDINTR;
1701 
1702 	head = ring->head;
1703 	while (count < budget) {
1704 		uint32_t prev_head = head;
1705 		struct mbuf *mhead, *mtail;
1706 		struct virtio_net_hdr *vh;
1707 		struct netmap_slot *slot;
1708 		unsigned int nmbuf_len;
1709 		uint8_t *nmbuf;
1710 		int deliver = 1; /* the mbuf to the network stack. */
1711 host_sync:
1712 		if (head == ring->tail) {
1713 			/* We ran out of slot, let's see if the host has
1714 			 * added some, by reading hwcur and hwtail from
1715 			 * the CSB. */
1716 			ptnet_sync_tail(ktoa, kring);
1717 
1718 			if (head == ring->tail) {
1719 				/* Still no slots available. Reactivate
1720 				 * interrupts as they were disabled by the
1721 				 * host thread right before issuing the
1722 				 * last interrupt. */
1723 				atok->appl_need_kick = 1;
1724 
1725 				/* Double check for more completed RX slots.
1726 				 * We need a full barrier to prevent the store
1727 				 * to atok->appl_need_kick to be reordered with
1728 				 * the load from ktoa->hwcur and ktoa->hwtail
1729 				 * (store-load barrier). */
1730 				nm_stld_barrier();
1731 				ptnet_sync_tail(ktoa, kring);
1732 				if (likely(head == ring->tail)) {
1733 					break;
1734 				}
1735 				atok->appl_need_kick = 0;
1736 			}
1737 		}
1738 
1739 		/* Initialize ring state variables, possibly grabbing the
1740 		 * virtio-net header. */
1741 		slot = ring->slot + head;
1742 		nmbuf = NMB(na, slot);
1743 		nmbuf_len = slot->len;
1744 
1745 		vh = (struct virtio_net_hdr *)nmbuf;
1746 		if (have_vnet_hdr) {
1747 			if (unlikely(nmbuf_len < PTNET_HDR_SIZE)) {
1748 				/* There is no good reason why host should
1749 				 * put the header in multiple netmap slots.
1750 				 * If this is the case, discard. */
1751 				nm_prlim(1, "Fragmented vnet-hdr: dropping");
1752 				head = ptnet_rx_discard(kring, head);
1753 				pq->stats.iqdrops ++;
1754 				deliver = 0;
1755 				goto skip;
1756 			}
1757 			nm_prdis(1, "%s: vnet hdr: flags %x csum_start %u "
1758 			      "csum_ofs %u hdr_len = %u gso_size %u "
1759 			      "gso_type %x", __func__, vh->flags,
1760 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
1761 			      vh->gso_size, vh->gso_type);
1762 			nmbuf += PTNET_HDR_SIZE;
1763 			nmbuf_len -= PTNET_HDR_SIZE;
1764 		}
1765 
1766 		/* Allocate the head of a new mbuf chain.
1767 		 * We use m_getcl() to allocate an mbuf with standard cluster
1768 		 * size (MCLBYTES). In the future we could use m_getjcl()
1769 		 * to choose different sizes. */
1770 		mhead = mtail = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1771 		if (unlikely(mhead == NULL)) {
1772 			device_printf(sc->dev, "%s: failed to allocate mbuf "
1773 				      "head\n", __func__);
1774 			pq->stats.errors ++;
1775 			break;
1776 		}
1777 
1778 		/* Initialize the mbuf state variables. */
1779 		mhead->m_pkthdr.len = nmbuf_len;
1780 		mtail->m_len = 0;
1781 
1782 		/* Scan all the netmap slots containing the current packet. */
1783 		for (;;) {
1784 			DBG(device_printf(sc->dev, "%s: h %u t %u rcv frag "
1785 					  "len %u, flags %u\n", __func__,
1786 					  head, ring->tail, slot->len,
1787 					  slot->flags));
1788 
1789 			mtail = ptnet_rx_slot(mtail, nmbuf, nmbuf_len);
1790 			if (unlikely(!mtail)) {
1791 				/* Ouch. We ran out of memory while processing
1792 				 * a packet. We have to restore the previous
1793 				 * head position, free the mbuf chain, and
1794 				 * schedule the taskqueue to give the packet
1795 				 * another chance. */
1796 				device_printf(sc->dev, "%s: failed to allocate"
1797 					" mbuf frag, reset head %u --> %u\n",
1798 					__func__, head, prev_head);
1799 				head = prev_head;
1800 				m_freem(mhead);
1801 				pq->stats.errors ++;
1802 				if (may_resched) {
1803 					taskqueue_enqueue(pq->taskq,
1804 							  &pq->task);
1805 				}
1806 				goto escape;
1807 			}
1808 
1809 			/* We have to increment head irrespective of the
1810 			 * NS_MOREFRAG being set or not. */
1811 			head = nm_next(head, lim);
1812 
1813 			if (!(slot->flags & NS_MOREFRAG)) {
1814 				break;
1815 			}
1816 
1817 			if (unlikely(head == ring->tail)) {
1818 				/* The very last slot prepared by the host has
1819 				 * the NS_MOREFRAG set. Drop it and continue
1820 				 * the outer cycle (to do the double-check). */
1821 				nm_prlim(1, "Incomplete packet: dropping");
1822 				m_freem(mhead);
1823 				pq->stats.iqdrops ++;
1824 				goto host_sync;
1825 			}
1826 
1827 			slot = ring->slot + head;
1828 			nmbuf = NMB(na, slot);
1829 			nmbuf_len = slot->len;
1830 			mhead->m_pkthdr.len += nmbuf_len;
1831 		}
1832 
1833 		mhead->m_pkthdr.rcvif = ifp;
1834 		mhead->m_pkthdr.csum_flags = 0;
1835 
1836 		/* Store the queue idx in the packet header. */
1837 		mhead->m_pkthdr.flowid = pq->kring_id;
1838 		M_HASHTYPE_SET(mhead, M_HASHTYPE_OPAQUE);
1839 
1840 		if (ifp->if_capenable & IFCAP_VLAN_HWTAGGING) {
1841 			struct ether_header *eh;
1842 
1843 			eh = mtod(mhead, struct ether_header *);
1844 			if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
1845 				ptnet_vlan_tag_remove(mhead);
1846 				/*
1847 				 * With the 802.1Q header removed, update the
1848 				 * checksum starting location accordingly.
1849 				 */
1850 				if (vh->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
1851 					vh->csum_start -= ETHER_VLAN_ENCAP_LEN;
1852 			}
1853 		}
1854 
1855 		if (unlikely(have_vnet_hdr && virtio_net_rx_csum(mhead, vh))) {
1856 			m_freem(mhead);
1857 			nm_prlim(1, "Csum offload error: dropping");
1858 			pq->stats.iqdrops ++;
1859 			deliver = 0;
1860 		}
1861 
1862 skip:
1863 		count ++;
1864 		if (++batch_count >= PTNET_RX_BATCH) {
1865 			/* Some packets have been (or will be) pushed to the network
1866 			 * stack. We need to update the CSB to tell the host about
1867 			 * the new ring->cur and ring->head (RX buffer refill). */
1868 			ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
1869 			batch_count = 0;
1870 		}
1871 
1872 		if (likely(deliver))  {
1873 			pq->stats.packets ++;
1874 			pq->stats.bytes += mhead->m_pkthdr.len;
1875 
1876 			PTNET_Q_UNLOCK(pq);
1877 			(*ifp->if_input)(ifp, mhead);
1878 			PTNET_Q_LOCK(pq);
1879 			/* The ring->head index (and related indices) are
1880 			 * updated under pq lock by ptnet_ring_update().
1881 			 * Since we dropped the lock to call if_input(), we
1882 			 * must reload ring->head and restart processing the
1883 			 * ring from there. */
1884 			head = ring->head;
1885 
1886 			if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1887 				/* The interface has gone down while we didn't
1888 				 * have the lock. Stop any processing and exit. */
1889 				goto unlock;
1890 			}
1891 		}
1892 	}
1893 escape:
1894 	if (batch_count) {
1895 		ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
1896 
1897 	}
1898 
1899 	if (count >= budget && may_resched) {
1900 		/* If we ran out of budget or the double-check found new
1901 		 * slots to process, schedule the taskqueue. */
1902 		DBG(nm_prlim(1, "out of budget: resched h %u t %u\n",
1903 					head, ring->tail));
1904 		taskqueue_enqueue(pq->taskq, &pq->task);
1905 	}
1906 unlock:
1907 	PTNET_Q_UNLOCK(pq);
1908 
1909 	return count;
1910 }
1911 
1912 static void
1913 ptnet_rx_task(void *context, int pending)
1914 {
1915 	struct ptnet_queue *pq = context;
1916 
1917 	DBG(nm_prlim(1, "%s: pq #%u\n", __func__, pq->kring_id));
1918 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
1919 }
1920 
1921 static void
1922 ptnet_tx_task(void *context, int pending)
1923 {
1924 	struct ptnet_queue *pq = context;
1925 
1926 	DBG(nm_prlim(1, "%s: pq #%u\n", __func__, pq->kring_id));
1927 	ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
1928 }
1929 
1930 #ifdef DEVICE_POLLING
1931 /* We don't need to handle differently POLL_AND_CHECK_STATUS and
1932  * POLL_ONLY, since we don't have an Interrupt Status Register. */
1933 static int
1934 ptnet_poll(if_t ifp, enum poll_cmd cmd, int budget)
1935 {
1936 	struct ptnet_softc *sc = if_getsoftc(ifp);
1937 	unsigned int queue_budget;
1938 	unsigned int count = 0;
1939 	bool borrow = false;
1940 	int i;
1941 
1942 	KASSERT(sc->num_rings > 0, ("Found no queues in while polling ptnet"));
1943 	queue_budget = MAX(budget / sc->num_rings, 1);
1944 	nm_prlim(1, "Per-queue budget is %d", queue_budget);
1945 
1946 	while (budget) {
1947 		unsigned int rcnt = 0;
1948 
1949 		for (i = 0; i < sc->num_rings; i++) {
1950 			struct ptnet_queue *pq = sc->queues + i;
1951 
1952 			if (borrow) {
1953 				queue_budget = MIN(queue_budget, budget);
1954 				if (queue_budget == 0) {
1955 					break;
1956 				}
1957 			}
1958 
1959 			if (i < sc->num_tx_rings) {
1960 				rcnt += ptnet_drain_transmit_queue(pq,
1961 						   queue_budget, false);
1962 			} else {
1963 				rcnt += ptnet_rx_eof(pq, queue_budget,
1964 						      false);
1965 			}
1966 		}
1967 
1968 		if (!rcnt) {
1969 			/* A scan of the queues gave no result, we can
1970 			 * stop here. */
1971 			break;
1972 		}
1973 
1974 		if (rcnt > budget) {
1975 			/* This may happen when initial budget < sc->num_rings,
1976 			 * since one packet budget is given to each queue
1977 			 * anyway. Just pretend we didn't eat "so much". */
1978 			rcnt = budget;
1979 		}
1980 		count += rcnt;
1981 		budget -= rcnt;
1982 		borrow = true;
1983 	}
1984 
1985 
1986 	return count;
1987 }
1988 #endif /* DEVICE_POLLING */
1989 #endif /* WITH_PTNETMAP */
1990