xref: /freebsd/sys/dev/netmap/if_ptnet.c (revision 3077532b1bb2911d3012ee90bae9d9499c960569)
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 
27 /* Driver for ptnet paravirtualized network device. */
28 
29 #include <sys/cdefs.h>
30 #include "opt_inet.h"
31 #include "opt_inet6.h"
32 
33 #include <sys/param.h>
34 #include <sys/systm.h>
35 #include <sys/kernel.h>
36 #include <sys/sockio.h>
37 #include <sys/mbuf.h>
38 #include <sys/malloc.h>
39 #include <sys/module.h>
40 #include <sys/socket.h>
41 #include <sys/sysctl.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/taskqueue.h>
45 #include <sys/smp.h>
46 #include <sys/time.h>
47 #include <machine/smp.h>
48 
49 #include <vm/uma.h>
50 #include <vm/vm.h>
51 #include <vm/pmap.h>
52 
53 #include <net/ethernet.h>
54 #include <net/if.h>
55 #include <net/if_var.h>
56 #include <net/if_arp.h>
57 #include <net/if_dl.h>
58 #include <net/if_types.h>
59 #include <net/if_media.h>
60 #include <net/if_vlan_var.h>
61 #include <net/bpf.h>
62 
63 #include <netinet/in_systm.h>
64 #include <netinet/in.h>
65 #include <netinet/ip.h>
66 #include <netinet/ip6.h>
67 #include <netinet6/ip6_var.h>
68 #include <netinet/udp.h>
69 #include <netinet/tcp.h>
70 
71 #include <machine/bus.h>
72 #include <machine/resource.h>
73 #include <sys/bus.h>
74 #include <sys/rman.h>
75 
76 #include <dev/pci/pcivar.h>
77 #include <dev/pci/pcireg.h>
78 
79 #include <sys/selinfo.h>
80 #include <net/netmap.h>
81 #include <dev/netmap/netmap_kern.h>
82 #include <net/netmap_virt.h>
83 #include <dev/netmap/netmap_mem2.h>
84 #include <dev/virtio/network/virtio_net.h>
85 
86 #ifdef WITH_PTNETMAP
87 
88 #ifndef INET
89 #error "INET not defined, cannot support offloadings"
90 #endif
91 
92 static uint64_t	ptnet_get_counter(if_t, ift_counter);
93 
94 //#define PTNETMAP_STATS
95 //#define DEBUG
96 #ifdef DEBUG
97 #define DBG(x) x
98 #else   /* !DEBUG */
99 #define DBG(x)
100 #endif  /* !DEBUG */
101 
102 extern int ptnet_vnet_hdr; /* Tunable parameter */
103 
104 struct ptnet_softc;
105 
106 struct ptnet_queue_stats {
107 	uint64_t	packets; /* if_[io]packets */
108 	uint64_t	bytes;	 /* if_[io]bytes */
109 	uint64_t	errors;	 /* if_[io]errors */
110 	uint64_t	iqdrops; /* if_iqdrops */
111 	uint64_t	mcasts;  /* if_[io]mcasts */
112 #ifdef PTNETMAP_STATS
113 	uint64_t	intrs;
114 	uint64_t	kicks;
115 #endif /* PTNETMAP_STATS */
116 };
117 
118 struct ptnet_queue {
119 	struct ptnet_softc		*sc;
120 	struct				resource *irq;
121 	void				*cookie;
122 	int				kring_id;
123 	struct nm_csb_atok		*atok;
124 	struct nm_csb_ktoa		*ktoa;
125 	unsigned int			kick;
126 	struct mtx			lock;
127 	struct buf_ring			*bufring; /* for TX queues */
128 	struct ptnet_queue_stats	stats;
129 #ifdef PTNETMAP_STATS
130 	struct ptnet_queue_stats	last_stats;
131 #endif /* PTNETMAP_STATS */
132 	struct taskqueue		*taskq;
133 	struct task			task;
134 	char				lock_name[16];
135 };
136 
137 #define PTNET_Q_LOCK(_pq)	mtx_lock(&(_pq)->lock)
138 #define PTNET_Q_TRYLOCK(_pq)	mtx_trylock(&(_pq)->lock)
139 #define PTNET_Q_UNLOCK(_pq)	mtx_unlock(&(_pq)->lock)
140 
141 struct ptnet_softc {
142 	device_t		dev;
143 	if_t			ifp;
144 	struct ifmedia		media;
145 	struct mtx		lock;
146 	char			lock_name[16];
147 	char			hwaddr[ETHER_ADDR_LEN];
148 
149 	/* Mirror of PTFEAT register. */
150 	uint32_t		ptfeatures;
151 	unsigned int		vnet_hdr_len;
152 
153 	/* PCI BARs support. */
154 	struct resource		*iomem;
155 	struct resource		*msix_mem;
156 
157 	unsigned int		num_rings;
158 	unsigned int		num_tx_rings;
159 	struct ptnet_queue	*queues;
160 	struct ptnet_queue	*rxqueues;
161 	struct nm_csb_atok	*csb_gh;
162 	struct nm_csb_ktoa	*csb_hg;
163 
164 	unsigned int		min_tx_space;
165 
166 	struct netmap_pt_guest_adapter *ptna;
167 
168 	struct callout		tick;
169 #ifdef PTNETMAP_STATS
170 	struct timeval		last_ts;
171 #endif /* PTNETMAP_STATS */
172 };
173 
174 #define PTNET_CORE_LOCK(_sc)	mtx_lock(&(_sc)->lock)
175 #define PTNET_CORE_UNLOCK(_sc)	mtx_unlock(&(_sc)->lock)
176 
177 static int	ptnet_probe(device_t);
178 static int	ptnet_attach(device_t);
179 static int	ptnet_detach(device_t);
180 static int	ptnet_suspend(device_t);
181 static int	ptnet_resume(device_t);
182 static int	ptnet_shutdown(device_t);
183 
184 static void	ptnet_init(void *opaque);
185 static int	ptnet_ioctl(if_t ifp, u_long cmd, caddr_t data);
186 static int	ptnet_init_locked(struct ptnet_softc *sc);
187 static int	ptnet_stop(struct ptnet_softc *sc);
188 static int	ptnet_transmit(if_t ifp, struct mbuf *m);
189 static int	ptnet_drain_transmit_queue(struct ptnet_queue *pq,
190 					   unsigned int budget,
191 					   bool may_resched);
192 static void	ptnet_qflush(if_t ifp);
193 static void	ptnet_tx_task(void *context, int pending);
194 
195 static int	ptnet_media_change(if_t ifp);
196 static void	ptnet_media_status(if_t ifp, struct ifmediareq *ifmr);
197 #ifdef PTNETMAP_STATS
198 static void	ptnet_tick(void *opaque);
199 #endif
200 
201 static int	ptnet_irqs_init(struct ptnet_softc *sc);
202 static void	ptnet_irqs_fini(struct ptnet_softc *sc);
203 
204 static uint32_t ptnet_nm_ptctl(struct ptnet_softc *sc, uint32_t cmd);
205 static int      ptnet_nm_config(struct netmap_adapter *na,
206 				struct nm_config_info *info);
207 static void	ptnet_update_vnet_hdr(struct ptnet_softc *sc);
208 static int	ptnet_nm_register(struct netmap_adapter *na, int onoff);
209 static int	ptnet_nm_txsync(struct netmap_kring *kring, int flags);
210 static int	ptnet_nm_rxsync(struct netmap_kring *kring, int flags);
211 static void	ptnet_nm_intr(struct netmap_adapter *na, int onoff);
212 
213 static void	ptnet_tx_intr(void *opaque);
214 static void	ptnet_rx_intr(void *opaque);
215 
216 static unsigned	ptnet_rx_discard(struct netmap_kring *kring,
217 				 unsigned int head);
218 static int	ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget,
219 			     bool may_resched);
220 static void	ptnet_rx_task(void *context, int pending);
221 
222 #ifdef DEVICE_POLLING
223 static poll_handler_t ptnet_poll;
224 #endif
225 
226 static device_method_t ptnet_methods[] = {
227 	DEVMETHOD(device_probe,			ptnet_probe),
228 	DEVMETHOD(device_attach,		ptnet_attach),
229 	DEVMETHOD(device_detach,		ptnet_detach),
230 	DEVMETHOD(device_suspend,		ptnet_suspend),
231 	DEVMETHOD(device_resume,		ptnet_resume),
232 	DEVMETHOD(device_shutdown,		ptnet_shutdown),
233 	DEVMETHOD_END
234 };
235 
236 static driver_t ptnet_driver = {
237 	"ptnet",
238 	ptnet_methods,
239 	sizeof(struct ptnet_softc)
240 };
241 
242 /* We use (SI_ORDER_MIDDLE+2) here, see DEV_MODULE_ORDERED() invocation. */
243 DRIVER_MODULE_ORDERED(ptnet, pci, ptnet_driver, NULL, NULL,
244 		      SI_ORDER_MIDDLE + 2);
245 
246 static int
ptnet_probe(device_t dev)247 ptnet_probe(device_t dev)
248 {
249 	if (pci_get_vendor(dev) != PTNETMAP_PCI_VENDOR_ID ||
250 		pci_get_device(dev) != PTNETMAP_PCI_NETIF_ID) {
251 		return (ENXIO);
252 	}
253 
254 	device_set_desc(dev, "ptnet network adapter");
255 
256 	return (BUS_PROBE_DEFAULT);
257 }
258 
ptnet_kick(struct ptnet_queue * pq)259 static inline void ptnet_kick(struct ptnet_queue *pq)
260 {
261 #ifdef PTNETMAP_STATS
262 	pq->stats.kicks ++;
263 #endif /* PTNETMAP_STATS */
264 	bus_write_4(pq->sc->iomem, pq->kick, 0);
265 }
266 
267 #define PTNET_BUF_RING_SIZE	4096
268 #define PTNET_RX_BUDGET		512
269 #define PTNET_RX_BATCH		1
270 #define PTNET_TX_BUDGET		512
271 #define PTNET_TX_BATCH		64
272 #define PTNET_HDR_SIZE		sizeof(struct virtio_net_hdr_mrg_rxbuf)
273 #define PTNET_MAX_PKT_SIZE	65536
274 
275 #define PTNET_CSUM_OFFLOAD	(CSUM_TCP | CSUM_UDP)
276 #define PTNET_CSUM_OFFLOAD_IPV6	(CSUM_TCP_IPV6 | CSUM_UDP_IPV6)
277 #define PTNET_ALL_OFFLOAD	(CSUM_TSO | PTNET_CSUM_OFFLOAD |\
278 				 PTNET_CSUM_OFFLOAD_IPV6)
279 
280 static int
ptnet_attach(device_t dev)281 ptnet_attach(device_t dev)
282 {
283 	uint32_t ptfeatures = 0;
284 	unsigned int num_rx_rings, num_tx_rings;
285 	struct netmap_adapter na_arg;
286 	unsigned int nifp_offset;
287 	struct ptnet_softc *sc;
288 	if_t ifp;
289 	uint32_t macreg;
290 	int err, rid;
291 	int i;
292 
293 	sc = device_get_softc(dev);
294 	sc->dev = dev;
295 
296 	/* Setup PCI resources. */
297 	pci_enable_busmaster(dev);
298 
299 	rid = PCIR_BAR(PTNETMAP_IO_PCI_BAR);
300 	sc->iomem = bus_alloc_resource_any(dev, SYS_RES_IOPORT, &rid,
301 					   RF_ACTIVE);
302 	if (sc->iomem == NULL) {
303 		device_printf(dev, "Failed to map I/O BAR\n");
304 		return (ENXIO);
305 	}
306 
307 	/* Negotiate features with the hypervisor. */
308 	if (ptnet_vnet_hdr) {
309 		ptfeatures |= PTNETMAP_F_VNET_HDR;
310 	}
311 	bus_write_4(sc->iomem, PTNET_IO_PTFEAT, ptfeatures); /* wanted */
312 	ptfeatures = bus_read_4(sc->iomem, PTNET_IO_PTFEAT); /* acked */
313 	sc->ptfeatures = ptfeatures;
314 
315 	num_tx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_RINGS);
316 	num_rx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_RINGS);
317 	sc->num_rings = num_tx_rings + num_rx_rings;
318 	sc->num_tx_rings = num_tx_rings;
319 
320 	if (sc->num_rings * sizeof(struct nm_csb_atok) > PAGE_SIZE) {
321 		device_printf(dev, "CSB cannot handle that many rings (%u)\n",
322 				sc->num_rings);
323 		err = ENOMEM;
324 		goto err_path;
325 	}
326 
327 	/* Allocate CSB and carry out CSB allocation protocol. */
328 	sc->csb_gh = contigmalloc(2*PAGE_SIZE, M_DEVBUF, M_NOWAIT | M_ZERO,
329 				  (size_t)0, -1UL, PAGE_SIZE, 0);
330 	if (sc->csb_gh == NULL) {
331 		device_printf(dev, "Failed to allocate CSB\n");
332 		err = ENOMEM;
333 		goto err_path;
334 	}
335 	sc->csb_hg = (struct nm_csb_ktoa *)(((char *)sc->csb_gh) + PAGE_SIZE);
336 
337 	{
338 		/*
339 		 * We use uint64_t rather than vm_paddr_t since we
340 		 * need 64 bit addresses even on 32 bit platforms.
341 		 */
342 		uint64_t paddr = vtophys(sc->csb_gh);
343 
344 		/* CSB allocation protocol: write to BAH first, then
345 		 * to BAL (for both GH and HG sections). */
346 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAH,
347 				(paddr >> 32) & 0xffffffff);
348 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAL,
349 				paddr & 0xffffffff);
350 		paddr = vtophys(sc->csb_hg);
351 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAH,
352 				(paddr >> 32) & 0xffffffff);
353 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAL,
354 				paddr & 0xffffffff);
355 	}
356 
357 	/* Allocate and initialize per-queue data structures. */
358 	sc->queues = malloc(sizeof(struct ptnet_queue) * sc->num_rings,
359 			    M_DEVBUF, M_NOWAIT | M_ZERO);
360 	if (sc->queues == NULL) {
361 		err = ENOMEM;
362 		goto err_path;
363 	}
364 	sc->rxqueues = sc->queues + num_tx_rings;
365 
366 	for (i = 0; i < sc->num_rings; i++) {
367 		struct ptnet_queue *pq = sc->queues + i;
368 
369 		pq->sc = sc;
370 		pq->kring_id = i;
371 		pq->kick = PTNET_IO_KICK_BASE + 4 * i;
372 		pq->atok = sc->csb_gh + i;
373 		pq->ktoa = sc->csb_hg + i;
374 		snprintf(pq->lock_name, sizeof(pq->lock_name), "%s-%d",
375 			 device_get_nameunit(dev), i);
376 		mtx_init(&pq->lock, pq->lock_name, NULL, MTX_DEF);
377 		if (i >= num_tx_rings) {
378 			/* RX queue: fix kring_id. */
379 			pq->kring_id -= num_tx_rings;
380 		} else {
381 			/* TX queue: allocate buf_ring. */
382 			pq->bufring = buf_ring_alloc(PTNET_BUF_RING_SIZE,
383 						M_DEVBUF, M_NOWAIT, &pq->lock);
384 			if (pq->bufring == NULL) {
385 				err = ENOMEM;
386 				goto err_path;
387 			}
388 		}
389 	}
390 
391 	sc->min_tx_space = 64; /* Safe initial value. */
392 
393 	err = ptnet_irqs_init(sc);
394 	if (err) {
395 		goto err_path;
396 	}
397 
398 	/* Setup Ethernet interface. */
399 	sc->ifp = ifp = if_alloc(IFT_ETHER);
400 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
401 	if_setbaudrate(ifp, IF_Gbps(10));
402 	if_setsoftc(ifp, sc);
403 	if_setflags(ifp, IFF_BROADCAST | IFF_MULTICAST | IFF_SIMPLEX);
404 	if_setinitfn(ifp, ptnet_init);
405 	if_setioctlfn(ifp, ptnet_ioctl);
406 	if_setget_counter(ifp, ptnet_get_counter);
407 	if_settransmitfn(ifp, ptnet_transmit);
408 	if_setqflushfn(ifp, ptnet_qflush);
409 
410 	ifmedia_init(&sc->media, IFM_IMASK, ptnet_media_change,
411 		     ptnet_media_status);
412 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T | IFM_FDX, 0, NULL);
413 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T | IFM_FDX);
414 
415 	macreg = bus_read_4(sc->iomem, PTNET_IO_MAC_HI);
416 	sc->hwaddr[0] = (macreg >> 8) & 0xff;
417 	sc->hwaddr[1] = macreg & 0xff;
418 	macreg = bus_read_4(sc->iomem, PTNET_IO_MAC_LO);
419 	sc->hwaddr[2] = (macreg >> 24) & 0xff;
420 	sc->hwaddr[3] = (macreg >> 16) & 0xff;
421 	sc->hwaddr[4] = (macreg >> 8) & 0xff;
422 	sc->hwaddr[5] = macreg & 0xff;
423 
424 	ether_ifattach(ifp, sc->hwaddr);
425 
426 	if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
427 	if_setcapabilitiesbit(ifp, IFCAP_JUMBO_MTU | IFCAP_VLAN_MTU, 0);
428 
429 	if (sc->ptfeatures & PTNETMAP_F_VNET_HDR) {
430 		/* Similarly to what the vtnet driver does, we can emulate
431 		 * VLAN offloadings by inserting and removing the 802.1Q
432 		 * header during transmit and receive. We are then able
433 		 * to do checksum offloading of VLAN frames. */
434 		if_setcapabilitiesbit(ifp, IFCAP_HWCSUM | IFCAP_HWCSUM_IPV6
435 					| IFCAP_VLAN_HWCSUM
436 					| IFCAP_TSO | IFCAP_LRO
437 					| IFCAP_VLAN_HWTSO
438 					| IFCAP_VLAN_HWTAGGING, 0);
439 	}
440 
441 	if_setcapenable(ifp, if_getcapabilities(ifp));
442 #ifdef DEVICE_POLLING
443 	/* Don't enable polling by default. */
444 	if_setcapabilitiesbit(ifp, IFCAP_POLLING, 0);
445 #endif
446 	snprintf(sc->lock_name, sizeof(sc->lock_name),
447 		 "%s", device_get_nameunit(dev));
448 	mtx_init(&sc->lock, sc->lock_name, "ptnet core lock", MTX_DEF);
449 	callout_init_mtx(&sc->tick, &sc->lock, 0);
450 
451 	/* Prepare a netmap_adapter struct instance to do netmap_attach(). */
452 	nifp_offset = bus_read_4(sc->iomem, PTNET_IO_NIFP_OFS);
453 	memset(&na_arg, 0, sizeof(na_arg));
454 	na_arg.ifp = ifp;
455 	na_arg.num_tx_desc = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_SLOTS);
456 	na_arg.num_rx_desc = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_SLOTS);
457 	na_arg.num_tx_rings = num_tx_rings;
458 	na_arg.num_rx_rings = num_rx_rings;
459 	na_arg.nm_config = ptnet_nm_config;
460 	na_arg.nm_krings_create = ptnet_nm_krings_create;
461 	na_arg.nm_krings_delete = ptnet_nm_krings_delete;
462 	na_arg.nm_dtor = ptnet_nm_dtor;
463 	na_arg.nm_intr = ptnet_nm_intr;
464 	na_arg.nm_register = ptnet_nm_register;
465 	na_arg.nm_txsync = ptnet_nm_txsync;
466 	na_arg.nm_rxsync = ptnet_nm_rxsync;
467 
468 	netmap_pt_guest_attach(&na_arg, nifp_offset,
469                                 bus_read_4(sc->iomem, PTNET_IO_HOSTMEMID));
470 
471 	/* Now a netmap adapter for this ifp has been allocated, and it
472 	 * can be accessed through NA(ifp). We also have to initialize the CSB
473 	 * pointer. */
474 	sc->ptna = (struct netmap_pt_guest_adapter *)NA(ifp);
475 
476 	/* If virtio-net header was negotiated, set the virt_hdr_len field in
477 	 * the netmap adapter, to inform users that this netmap adapter requires
478 	 * the application to deal with the headers. */
479 	ptnet_update_vnet_hdr(sc);
480 
481 	device_printf(dev, "%s() completed\n", __func__);
482 
483 	return (0);
484 
485 err_path:
486 	ptnet_detach(dev);
487 	return err;
488 }
489 
490 /* Stop host sync-kloop if it was running. */
491 static void
ptnet_device_shutdown(struct ptnet_softc * sc)492 ptnet_device_shutdown(struct ptnet_softc *sc)
493 {
494 	ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_DELETE);
495 	bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAH, 0);
496 	bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAL, 0);
497 	bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAH, 0);
498 	bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAL, 0);
499 }
500 
501 static int
ptnet_detach(device_t dev)502 ptnet_detach(device_t dev)
503 {
504 	struct ptnet_softc *sc = device_get_softc(dev);
505 	int i;
506 
507 	ptnet_device_shutdown(sc);
508 
509 #ifdef DEVICE_POLLING
510 	if (if_getcapenable(sc->ifp) & IFCAP_POLLING) {
511 		ether_poll_deregister(sc->ifp);
512 	}
513 #endif
514 	callout_drain(&sc->tick);
515 
516 	if (sc->queues) {
517 		/* Drain taskqueues before calling if_detach. */
518 		for (i = 0; i < sc->num_rings; i++) {
519 			struct ptnet_queue *pq = sc->queues + i;
520 
521 			if (pq->taskq) {
522 				taskqueue_drain(pq->taskq, &pq->task);
523 			}
524 		}
525 	}
526 
527 	if (sc->ifp) {
528 		ether_ifdetach(sc->ifp);
529 
530 		/* Uninitialize netmap adapters for this device. */
531 		netmap_detach(sc->ifp);
532 
533 		ifmedia_removeall(&sc->media);
534 		if_free(sc->ifp);
535 		sc->ifp = NULL;
536 	}
537 
538 	ptnet_irqs_fini(sc);
539 
540 	if (sc->csb_gh) {
541 		free(sc->csb_gh, M_DEVBUF);
542 		sc->csb_gh = NULL;
543 		sc->csb_hg = NULL;
544 	}
545 
546 	if (sc->queues) {
547 		for (i = 0; i < sc->num_rings; i++) {
548 			struct ptnet_queue *pq = sc->queues + i;
549 
550 			if (mtx_initialized(&pq->lock)) {
551 				mtx_destroy(&pq->lock);
552 			}
553 			if (pq->bufring != NULL) {
554 				buf_ring_free(pq->bufring, M_DEVBUF);
555 			}
556 		}
557 		free(sc->queues, M_DEVBUF);
558 		sc->queues = NULL;
559 	}
560 
561 	if (sc->iomem) {
562 		bus_release_resource(dev, SYS_RES_IOPORT,
563 				     PCIR_BAR(PTNETMAP_IO_PCI_BAR), sc->iomem);
564 		sc->iomem = NULL;
565 	}
566 
567 	mtx_destroy(&sc->lock);
568 
569 	device_printf(dev, "%s() completed\n", __func__);
570 
571 	return (0);
572 }
573 
574 static int
ptnet_suspend(device_t dev)575 ptnet_suspend(device_t dev)
576 {
577 	struct ptnet_softc *sc = device_get_softc(dev);
578 
579 	(void)sc;
580 
581 	return (0);
582 }
583 
584 static int
ptnet_resume(device_t dev)585 ptnet_resume(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
ptnet_shutdown(device_t dev)595 ptnet_shutdown(device_t dev)
596 {
597 	struct ptnet_softc *sc = device_get_softc(dev);
598 
599 	ptnet_device_shutdown(sc);
600 
601 	return (0);
602 }
603 
604 static int
ptnet_irqs_init(struct ptnet_softc * sc)605 ptnet_irqs_init(struct ptnet_softc *sc)
606 {
607 	int rid = PCIR_BAR(PTNETMAP_MSIX_PCI_BAR);
608 	int nvecs = sc->num_rings;
609 	device_t dev = sc->dev;
610 	int err = ENOSPC;
611 	int cpu_cur;
612 	int i;
613 
614 	if (pci_find_cap(dev, PCIY_MSIX, NULL) != 0)  {
615 		device_printf(dev, "Could not find MSI-X capability\n");
616 		return (ENXIO);
617 	}
618 
619 	sc->msix_mem = bus_alloc_resource_any(dev, SYS_RES_MEMORY,
620 					      &rid, RF_ACTIVE);
621 	if (sc->msix_mem == NULL) {
622 		device_printf(dev, "Failed to allocate MSIX PCI BAR\n");
623 		return (ENXIO);
624 	}
625 
626 	if (pci_msix_count(dev) < nvecs) {
627 		device_printf(dev, "Not enough MSI-X vectors\n");
628 		goto err_path;
629 	}
630 
631 	err = pci_alloc_msix(dev, &nvecs);
632 	if (err) {
633 		device_printf(dev, "Failed to allocate MSI-X vectors\n");
634 		goto err_path;
635 	}
636 
637 	for (i = 0; i < nvecs; i++) {
638 		struct ptnet_queue *pq = sc->queues + i;
639 
640 		rid = i + 1;
641 		pq->irq = bus_alloc_resource_any(dev, SYS_RES_IRQ, &rid,
642 						 RF_ACTIVE);
643 		if (pq->irq == NULL) {
644 			device_printf(dev, "Failed to allocate interrupt "
645 					   "for queue #%d\n", i);
646 			err = ENOSPC;
647 			goto err_path;
648 		}
649 	}
650 
651 	cpu_cur = CPU_FIRST();
652 	for (i = 0; i < nvecs; i++) {
653 		struct ptnet_queue *pq = sc->queues + i;
654 		void (*handler)(void *) = ptnet_tx_intr;
655 
656 		if (i >= sc->num_tx_rings) {
657 			handler = ptnet_rx_intr;
658 		}
659 		err = bus_setup_intr(dev, pq->irq, INTR_TYPE_NET | INTR_MPSAFE,
660 				     NULL /* intr_filter */, handler,
661 				     pq, &pq->cookie);
662 		if (err) {
663 			device_printf(dev, "Failed to register intr handler "
664 					   "for queue #%d\n", i);
665 			goto err_path;
666 		}
667 
668 		bus_describe_intr(dev, pq->irq, pq->cookie, "q%d", i);
669 #if 0
670 		bus_bind_intr(sc->dev, pq->irq, cpu_cur);
671 #endif
672 		cpu_cur = CPU_NEXT(cpu_cur);
673 	}
674 
675 	device_printf(dev, "Allocated %d MSI-X vectors\n", nvecs);
676 
677 	cpu_cur = CPU_FIRST();
678 	for (i = 0; i < nvecs; i++) {
679 		struct ptnet_queue *pq = sc->queues + i;
680 
681 		if (i < sc->num_tx_rings)
682 			TASK_INIT(&pq->task, 0, ptnet_tx_task, pq);
683 		else
684 			NET_TASK_INIT(&pq->task, 0, ptnet_rx_task, pq);
685 
686 		pq->taskq = taskqueue_create_fast("ptnet_queue", M_NOWAIT,
687 					taskqueue_thread_enqueue, &pq->taskq);
688 		taskqueue_start_threads(&pq->taskq, 1, PI_NET, "%s-pq-%d",
689 					device_get_nameunit(sc->dev), cpu_cur);
690 		cpu_cur = CPU_NEXT(cpu_cur);
691 	}
692 
693 	return 0;
694 err_path:
695 	ptnet_irqs_fini(sc);
696 	return err;
697 }
698 
699 static void
ptnet_irqs_fini(struct ptnet_softc * sc)700 ptnet_irqs_fini(struct ptnet_softc *sc)
701 {
702 	device_t dev = sc->dev;
703 	int i;
704 
705 	for (i = 0; i < sc->num_rings; i++) {
706 		struct ptnet_queue *pq = sc->queues + i;
707 
708 		if (pq->taskq) {
709 			taskqueue_free(pq->taskq);
710 			pq->taskq = NULL;
711 		}
712 
713 		if (pq->cookie) {
714 			bus_teardown_intr(dev, pq->irq, pq->cookie);
715 			pq->cookie = NULL;
716 		}
717 
718 		if (pq->irq) {
719 			bus_release_resource(dev, SYS_RES_IRQ, i + 1, pq->irq);
720 			pq->irq = NULL;
721 		}
722 	}
723 
724 	if (sc->msix_mem) {
725 		pci_release_msi(dev);
726 
727 		bus_release_resource(dev, SYS_RES_MEMORY,
728 				     PCIR_BAR(PTNETMAP_MSIX_PCI_BAR),
729 				     sc->msix_mem);
730 		sc->msix_mem = NULL;
731 	}
732 }
733 
734 static void
ptnet_init(void * opaque)735 ptnet_init(void *opaque)
736 {
737 	struct ptnet_softc *sc = opaque;
738 
739 	PTNET_CORE_LOCK(sc);
740 	ptnet_init_locked(sc);
741 	PTNET_CORE_UNLOCK(sc);
742 }
743 
744 static int
ptnet_ioctl(if_t ifp,u_long cmd,caddr_t data)745 ptnet_ioctl(if_t ifp, u_long cmd, caddr_t data)
746 {
747 	struct ptnet_softc *sc = if_getsoftc(ifp);
748 	device_t dev = sc->dev;
749 	struct ifreq *ifr = (struct ifreq *)data;
750 	int mask __unused, err = 0;
751 
752 	switch (cmd) {
753 	case SIOCSIFFLAGS:
754 		device_printf(dev, "SIOCSIFFLAGS %x\n", if_getflags(ifp));
755 		PTNET_CORE_LOCK(sc);
756 		if (if_getflags(ifp) & IFF_UP) {
757 			/* Network stack wants the iff to be up. */
758 			err = ptnet_init_locked(sc);
759 		} else {
760 			/* Network stack wants the iff to be down. */
761 			err = ptnet_stop(sc);
762 		}
763 		/* We don't need to do nothing to support IFF_PROMISC,
764 		 * since that is managed by the backend port. */
765 		PTNET_CORE_UNLOCK(sc);
766 		break;
767 
768 	case SIOCSIFCAP:
769 		device_printf(dev, "SIOCSIFCAP %x %x\n",
770 			      ifr->ifr_reqcap, if_getcapenable(ifp));
771 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
772 #ifdef DEVICE_POLLING
773 		if (mask & IFCAP_POLLING) {
774 			struct ptnet_queue *pq;
775 			int i;
776 
777 			if (ifr->ifr_reqcap & IFCAP_POLLING) {
778 				err = ether_poll_register(ptnet_poll, ifp);
779 				if (err) {
780 					break;
781 				}
782 				/* Stop queues and sync with taskqueues. */
783 				if_setdrvflagbits(ifp, 0, IFF_DRV_RUNNING);
784 				for (i = 0; i < sc->num_rings; i++) {
785 					pq = sc-> queues + i;
786 					/* Make sure the worker sees the
787 					 * IFF_DRV_RUNNING down. */
788 					PTNET_Q_LOCK(pq);
789 					pq->atok->appl_need_kick = 0;
790 					PTNET_Q_UNLOCK(pq);
791 					/* Wait for rescheduling to finish. */
792 					if (pq->taskq) {
793 						taskqueue_drain(pq->taskq,
794 								&pq->task);
795 					}
796 				}
797 				if_setdrvflagbits(ifp, IFF_DRV_RUNNING, 0);
798 			} else {
799 				err = ether_poll_deregister(ifp);
800 				for (i = 0; i < sc->num_rings; i++) {
801 					pq = sc-> queues + i;
802 					PTNET_Q_LOCK(pq);
803 					pq->atok->appl_need_kick = 1;
804 					PTNET_Q_UNLOCK(pq);
805 				}
806 			}
807 		}
808 #endif  /* DEVICE_POLLING */
809 		if_setcapenable(ifp, ifr->ifr_reqcap);
810 		break;
811 
812 	case SIOCSIFMTU:
813 		/* We support any reasonable MTU. */
814 		if (ifr->ifr_mtu < ETHERMIN ||
815 				ifr->ifr_mtu > PTNET_MAX_PKT_SIZE) {
816 			err = EINVAL;
817 		} else {
818 			PTNET_CORE_LOCK(sc);
819 			if_setmtu(ifp, ifr->ifr_mtu);
820 			PTNET_CORE_UNLOCK(sc);
821 		}
822 		break;
823 
824 	case SIOCSIFMEDIA:
825 	case SIOCGIFMEDIA:
826 		err = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
827 		break;
828 
829 	default:
830 		err = ether_ioctl(ifp, cmd, data);
831 		break;
832 	}
833 
834 	return err;
835 }
836 
837 static int
ptnet_init_locked(struct ptnet_softc * sc)838 ptnet_init_locked(struct ptnet_softc *sc)
839 {
840 	if_t ifp = sc->ifp;
841 	struct netmap_adapter *na_dr = &sc->ptna->dr.up;
842 	struct netmap_adapter *na_nm = &sc->ptna->hwup.up;
843 	unsigned int nm_buf_size;
844 	int ret;
845 
846 	if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
847 		return 0; /* nothing to do */
848 	}
849 
850 	device_printf(sc->dev, "%s\n", __func__);
851 
852 	/* Translate offload capabilities according to if_capenable. */
853 	if_sethwassist(ifp, 0);
854 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
855 		if_sethwassistbits(ifp, PTNET_CSUM_OFFLOAD, 0);
856 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
857 		if_sethwassistbits(ifp, PTNET_CSUM_OFFLOAD_IPV6, 0);
858 	if (if_getcapenable(ifp) & IFCAP_TSO4)
859 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
860 	if (if_getcapenable(ifp) & IFCAP_TSO6)
861 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
862 
863 	/*
864 	 * Prepare the interface for netmap mode access.
865 	 */
866 	netmap_update_config(na_dr);
867 
868 	ret = netmap_mem_finalize(na_dr->nm_mem, na_dr);
869 	if (ret) {
870 		device_printf(sc->dev, "netmap_mem_finalize() failed\n");
871 		return ret;
872 	}
873 
874 	if (sc->ptna->backend_users == 0) {
875 		ret = ptnet_nm_krings_create(na_nm);
876 		if (ret) {
877 			device_printf(sc->dev, "ptnet_nm_krings_create() "
878 					       "failed\n");
879 			goto err_mem_finalize;
880 		}
881 
882 		ret = netmap_mem_rings_create(na_dr);
883 		if (ret) {
884 			device_printf(sc->dev, "netmap_mem_rings_create() "
885 					       "failed\n");
886 			goto err_rings_create;
887 		}
888 
889 		ret = netmap_mem_get_lut(na_dr->nm_mem, &na_dr->na_lut);
890 		if (ret) {
891 			device_printf(sc->dev, "netmap_mem_get_lut() "
892 					       "failed\n");
893 			goto err_get_lut;
894 		}
895 	}
896 
897 	ret = ptnet_nm_register(na_dr, 1 /* on */);
898 	if (ret) {
899 		goto err_register;
900 	}
901 
902 	nm_buf_size = NETMAP_BUF_SIZE(na_dr);
903 
904 	KASSERT(nm_buf_size > 0, ("Invalid netmap buffer size"));
905 	sc->min_tx_space = PTNET_MAX_PKT_SIZE / nm_buf_size + 2;
906 	device_printf(sc->dev, "%s: min_tx_space = %u\n", __func__,
907 		      sc->min_tx_space);
908 #ifdef PTNETMAP_STATS
909 	callout_reset(&sc->tick, hz, ptnet_tick, sc);
910 #endif
911 
912 	if_setdrvflagbits(ifp, IFF_DRV_RUNNING, 0);
913 
914 	return 0;
915 
916 err_register:
917 	memset(&na_dr->na_lut, 0, sizeof(na_dr->na_lut));
918 err_get_lut:
919 	netmap_mem_rings_delete(na_dr);
920 err_rings_create:
921 	ptnet_nm_krings_delete(na_nm);
922 err_mem_finalize:
923 	netmap_mem_deref(na_dr->nm_mem, na_dr);
924 
925 	return ret;
926 }
927 
928 /* To be called under core lock. */
929 static int
ptnet_stop(struct ptnet_softc * sc)930 ptnet_stop(struct ptnet_softc *sc)
931 {
932 	if_t ifp = sc->ifp;
933 	struct netmap_adapter *na_dr = &sc->ptna->dr.up;
934 	struct netmap_adapter *na_nm = &sc->ptna->hwup.up;
935 	int i;
936 
937 	device_printf(sc->dev, "%s\n", __func__);
938 
939 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)) {
940 		return 0; /* nothing to do */
941 	}
942 
943 	/* Clear the driver-ready flag, and synchronize with all the queues,
944 	 * so that after this loop we are sure nobody is working anymore with
945 	 * the device. This scheme is taken from the vtnet driver. */
946 	if_setdrvflagbits(ifp, 0, IFF_DRV_RUNNING);
947 	callout_stop(&sc->tick);
948 	for (i = 0; i < sc->num_rings; i++) {
949 		PTNET_Q_LOCK(sc->queues + i);
950 		PTNET_Q_UNLOCK(sc->queues + i);
951 	}
952 
953 	ptnet_nm_register(na_dr, 0 /* off */);
954 
955 	if (sc->ptna->backend_users == 0) {
956 		netmap_mem_rings_delete(na_dr);
957 		ptnet_nm_krings_delete(na_nm);
958 	}
959 	netmap_mem_deref(na_dr->nm_mem, na_dr);
960 
961 	return 0;
962 }
963 
964 static void
ptnet_qflush(if_t ifp)965 ptnet_qflush(if_t ifp)
966 {
967 	struct ptnet_softc *sc = if_getsoftc(ifp);
968 	int i;
969 
970 	/* Flush all the bufrings and do the interface flush. */
971 	for (i = 0; i < sc->num_rings; i++) {
972 		struct ptnet_queue *pq = sc->queues + i;
973 		struct mbuf *m;
974 
975 		PTNET_Q_LOCK(pq);
976 		if (pq->bufring) {
977 			while ((m = buf_ring_dequeue_sc(pq->bufring))) {
978 				m_freem(m);
979 			}
980 		}
981 		PTNET_Q_UNLOCK(pq);
982 	}
983 
984 	if_qflush(ifp);
985 }
986 
987 static int
ptnet_media_change(if_t ifp)988 ptnet_media_change(if_t ifp)
989 {
990 	struct ptnet_softc *sc = if_getsoftc(ifp);
991 	struct ifmedia *ifm = &sc->media;
992 
993 	if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER) {
994 		return EINVAL;
995 	}
996 
997 	return 0;
998 }
999 
1000 static uint64_t
ptnet_get_counter(if_t ifp,ift_counter cnt)1001 ptnet_get_counter(if_t ifp, ift_counter cnt)
1002 {
1003 	struct ptnet_softc *sc = if_getsoftc(ifp);
1004 	struct ptnet_queue_stats stats[2];
1005 	int i;
1006 
1007 	/* Accumulate statistics over the queues. */
1008 	memset(stats, 0, sizeof(stats));
1009 	for (i = 0; i < sc->num_rings; i++) {
1010 		struct ptnet_queue *pq = sc->queues + i;
1011 		int idx = (i < sc->num_tx_rings) ? 0 : 1;
1012 
1013 		stats[idx].packets	+= pq->stats.packets;
1014 		stats[idx].bytes	+= pq->stats.bytes;
1015 		stats[idx].errors	+= pq->stats.errors;
1016 		stats[idx].iqdrops	+= pq->stats.iqdrops;
1017 		stats[idx].mcasts	+= pq->stats.mcasts;
1018 	}
1019 
1020 	switch (cnt) {
1021 	case IFCOUNTER_IPACKETS:
1022 		return (stats[1].packets);
1023 	case IFCOUNTER_IQDROPS:
1024 		return (stats[1].iqdrops);
1025 	case IFCOUNTER_IERRORS:
1026 		return (stats[1].errors);
1027 	case IFCOUNTER_OPACKETS:
1028 		return (stats[0].packets);
1029 	case IFCOUNTER_OBYTES:
1030 		return (stats[0].bytes);
1031 	case IFCOUNTER_OMCASTS:
1032 		return (stats[0].mcasts);
1033 	default:
1034 		return (if_get_counter_default(ifp, cnt));
1035 	}
1036 }
1037 
1038 
1039 #ifdef PTNETMAP_STATS
1040 /* Called under core lock. */
1041 static void
ptnet_tick(void * opaque)1042 ptnet_tick(void *opaque)
1043 {
1044 	struct ptnet_softc *sc = opaque;
1045 	int i;
1046 
1047 	for (i = 0; i < sc->num_rings; i++) {
1048 		struct ptnet_queue *pq = sc->queues + i;
1049 		struct ptnet_queue_stats cur = pq->stats;
1050 		struct timeval now;
1051 		unsigned int delta;
1052 
1053 		microtime(&now);
1054 		delta = now.tv_usec - sc->last_ts.tv_usec +
1055 			(now.tv_sec - sc->last_ts.tv_sec) * 1000000;
1056 		delta /= 1000; /* in milliseconds */
1057 
1058 		if (delta == 0)
1059 			continue;
1060 
1061 		device_printf(sc->dev, "#%d[%u ms]:pkts %lu, kicks %lu, "
1062 			      "intr %lu\n", i, delta,
1063 			      (cur.packets - pq->last_stats.packets),
1064 			      (cur.kicks - pq->last_stats.kicks),
1065 			      (cur.intrs - pq->last_stats.intrs));
1066 		pq->last_stats = cur;
1067 	}
1068 	microtime(&sc->last_ts);
1069 	callout_schedule(&sc->tick, hz);
1070 }
1071 #endif /* PTNETMAP_STATS */
1072 
1073 static void
ptnet_media_status(if_t ifp,struct ifmediareq * ifmr)1074 ptnet_media_status(if_t ifp, struct ifmediareq *ifmr)
1075 {
1076 	/* We are always active, as the backend netmap port is
1077 	 * always open in netmap mode. */
1078 	ifmr->ifm_status = IFM_AVALID | IFM_ACTIVE;
1079 	ifmr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
1080 }
1081 
1082 static uint32_t
ptnet_nm_ptctl(struct ptnet_softc * sc,uint32_t cmd)1083 ptnet_nm_ptctl(struct ptnet_softc *sc, uint32_t cmd)
1084 {
1085 	/*
1086 	 * Write a command and read back error status,
1087 	 * with zero meaning success.
1088 	 */
1089 	bus_write_4(sc->iomem, PTNET_IO_PTCTL, cmd);
1090 	return bus_read_4(sc->iomem, PTNET_IO_PTCTL);
1091 }
1092 
1093 static int
ptnet_nm_config(struct netmap_adapter * na,struct nm_config_info * info)1094 ptnet_nm_config(struct netmap_adapter *na, struct nm_config_info *info)
1095 {
1096 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1097 
1098 	info->num_tx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_RINGS);
1099 	info->num_rx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_RINGS);
1100 	info->num_tx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_SLOTS);
1101 	info->num_rx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_SLOTS);
1102 	info->rx_buf_maxsize = NETMAP_BUF_SIZE(na);
1103 
1104 	device_printf(sc->dev, "txr %u, rxr %u, txd %u, rxd %u, rxbufsz %u\n",
1105 			info->num_tx_rings, info->num_rx_rings,
1106 			info->num_tx_descs, info->num_rx_descs,
1107 			info->rx_buf_maxsize);
1108 
1109 	return 0;
1110 }
1111 
1112 static void
ptnet_sync_from_csb(struct ptnet_softc * sc,struct netmap_adapter * na)1113 ptnet_sync_from_csb(struct ptnet_softc *sc, struct netmap_adapter *na)
1114 {
1115 	int i;
1116 
1117 	/* Sync krings from the host, reading from
1118 	 * CSB. */
1119 	for (i = 0; i < sc->num_rings; i++) {
1120 		struct nm_csb_atok *atok = sc->queues[i].atok;
1121 		struct nm_csb_ktoa *ktoa = sc->queues[i].ktoa;
1122 		struct netmap_kring *kring;
1123 
1124 		if (i < na->num_tx_rings) {
1125 			kring = na->tx_rings[i];
1126 		} else {
1127 			kring = na->rx_rings[i - na->num_tx_rings];
1128 		}
1129 		kring->rhead = kring->ring->head = atok->head;
1130 		kring->rcur = kring->ring->cur = atok->cur;
1131 		kring->nr_hwcur = ktoa->hwcur;
1132 		kring->nr_hwtail = kring->rtail =
1133 			kring->ring->tail = ktoa->hwtail;
1134 
1135 		nm_prdis("%d,%d: csb {hc %u h %u c %u ht %u}", t, i,
1136 		   ktoa->hwcur, atok->head, atok->cur,
1137 		   ktoa->hwtail);
1138 		nm_prdis("%d,%d: kring {hc %u rh %u rc %u h %u c %u ht %u rt %u t %u}",
1139 		   t, i, kring->nr_hwcur, kring->rhead, kring->rcur,
1140 		   kring->ring->head, kring->ring->cur, kring->nr_hwtail,
1141 		   kring->rtail, kring->ring->tail);
1142 	}
1143 }
1144 
1145 static void
ptnet_update_vnet_hdr(struct ptnet_softc * sc)1146 ptnet_update_vnet_hdr(struct ptnet_softc *sc)
1147 {
1148 	unsigned int wanted_hdr_len = ptnet_vnet_hdr ? PTNET_HDR_SIZE : 0;
1149 
1150 	bus_write_4(sc->iomem, PTNET_IO_VNET_HDR_LEN, wanted_hdr_len);
1151 	sc->vnet_hdr_len = bus_read_4(sc->iomem, PTNET_IO_VNET_HDR_LEN);
1152 	sc->ptna->hwup.up.virt_hdr_len = sc->vnet_hdr_len;
1153 }
1154 
1155 static int
ptnet_nm_register(struct netmap_adapter * na,int onoff)1156 ptnet_nm_register(struct netmap_adapter *na, int onoff)
1157 {
1158 	/* device-specific */
1159 	if_t ifp = na->ifp;
1160 	struct ptnet_softc *sc = if_getsoftc(ifp);
1161 	int native = (na == &sc->ptna->hwup.up);
1162 	struct ptnet_queue *pq;
1163 	int ret = 0;
1164 	int i;
1165 
1166 	if (!onoff) {
1167 		sc->ptna->backend_users--;
1168 	}
1169 
1170 	/* If this is the last netmap client, guest interrupt enable flags may
1171 	 * be in arbitrary state. Since these flags are going to be used also
1172 	 * by the netdevice driver, we have to make sure to start with
1173 	 * notifications enabled. Also, schedule NAPI to flush pending packets
1174 	 * in the RX rings, since we will not receive further interrupts
1175 	 * until these will be processed. */
1176 	if (native && !onoff && na->active_fds == 0) {
1177 		nm_prinf("Exit netmap mode, re-enable interrupts");
1178 		for (i = 0; i < sc->num_rings; i++) {
1179 			pq = sc->queues + i;
1180 			pq->atok->appl_need_kick = 1;
1181 		}
1182 	}
1183 
1184 	if (onoff) {
1185 		if (sc->ptna->backend_users == 0) {
1186 			/* Initialize notification enable fields in the CSB. */
1187 			for (i = 0; i < sc->num_rings; i++) {
1188 				pq = sc->queues + i;
1189 				pq->ktoa->kern_need_kick = 1;
1190 				pq->atok->appl_need_kick =
1191 					(!(if_getcapenable(ifp) & IFCAP_POLLING)
1192 						&& i >= sc->num_tx_rings);
1193 			}
1194 
1195 			/* Set the virtio-net header length. */
1196 			ptnet_update_vnet_hdr(sc);
1197 
1198 			/* Make sure the host adapter passed through is ready
1199 			 * for txsync/rxsync. */
1200 			ret = ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_CREATE);
1201 			if (ret) {
1202 				return ret;
1203 			}
1204 
1205 			/* Align the guest krings and rings to the state stored
1206 			 * in the CSB. */
1207 			ptnet_sync_from_csb(sc, na);
1208 		}
1209 
1210 		/* If not native, don't call nm_set_native_flags, since we don't want
1211 		 * to replace if_transmit method, nor set NAF_NETMAP_ON */
1212 		if (native) {
1213 			netmap_krings_mode_commit(na, onoff);
1214 			nm_set_native_flags(na);
1215 		}
1216 
1217 	} else {
1218 		if (native) {
1219 			nm_clear_native_flags(na);
1220 			netmap_krings_mode_commit(na, onoff);
1221 		}
1222 
1223 		if (sc->ptna->backend_users == 0) {
1224 			ret = ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_DELETE);
1225 		}
1226 	}
1227 
1228 	if (onoff) {
1229 		sc->ptna->backend_users++;
1230 	}
1231 
1232 	return ret;
1233 }
1234 
1235 static int
ptnet_nm_txsync(struct netmap_kring * kring,int flags)1236 ptnet_nm_txsync(struct netmap_kring *kring, int flags)
1237 {
1238 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1239 	struct ptnet_queue *pq = sc->queues + kring->ring_id;
1240 	bool notify;
1241 
1242 	notify = netmap_pt_guest_txsync(pq->atok, pq->ktoa, kring, flags);
1243 	if (notify) {
1244 		ptnet_kick(pq);
1245 	}
1246 
1247 	return 0;
1248 }
1249 
1250 static int
ptnet_nm_rxsync(struct netmap_kring * kring,int flags)1251 ptnet_nm_rxsync(struct netmap_kring *kring, int flags)
1252 {
1253 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1254 	struct ptnet_queue *pq = sc->rxqueues + kring->ring_id;
1255 	bool notify;
1256 
1257 	notify = netmap_pt_guest_rxsync(pq->atok, pq->ktoa, kring, flags);
1258 	if (notify) {
1259 		ptnet_kick(pq);
1260 	}
1261 
1262 	return 0;
1263 }
1264 
1265 static void
ptnet_nm_intr(struct netmap_adapter * na,int onoff)1266 ptnet_nm_intr(struct netmap_adapter *na, int onoff)
1267 {
1268 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1269 	int i;
1270 
1271 	for (i = 0; i < sc->num_rings; i++) {
1272 		struct ptnet_queue *pq = sc->queues + i;
1273 		pq->atok->appl_need_kick = onoff;
1274 	}
1275 }
1276 
1277 static void
ptnet_tx_intr(void * opaque)1278 ptnet_tx_intr(void *opaque)
1279 {
1280 	struct ptnet_queue *pq = opaque;
1281 	struct ptnet_softc *sc = pq->sc;
1282 
1283 	DBG(device_printf(sc->dev, "Tx interrupt #%d\n", pq->kring_id));
1284 #ifdef PTNETMAP_STATS
1285 	pq->stats.intrs ++;
1286 #endif /* PTNETMAP_STATS */
1287 
1288 	if (netmap_tx_irq(sc->ifp, pq->kring_id) != NM_IRQ_PASS) {
1289 		return;
1290 	}
1291 
1292 	/* Schedule the tasqueue to flush process transmissions requests.
1293 	 * However, vtnet, if_em and if_igb just call ptnet_transmit() here,
1294 	 * at least when using MSI-X interrupts. The if_em driver, instead
1295 	 * schedule taskqueue when using legacy interrupts. */
1296 	taskqueue_enqueue(pq->taskq, &pq->task);
1297 }
1298 
1299 static void
ptnet_rx_intr(void * opaque)1300 ptnet_rx_intr(void *opaque)
1301 {
1302 	struct ptnet_queue *pq = opaque;
1303 	struct ptnet_softc *sc = pq->sc;
1304 	unsigned int unused;
1305 
1306 	DBG(device_printf(sc->dev, "Rx interrupt #%d\n", pq->kring_id));
1307 #ifdef PTNETMAP_STATS
1308 	pq->stats.intrs ++;
1309 #endif /* PTNETMAP_STATS */
1310 
1311 	if (netmap_rx_irq(sc->ifp, pq->kring_id, &unused) != NM_IRQ_PASS) {
1312 		return;
1313 	}
1314 
1315 	/* Like vtnet, if_igb and if_em drivers when using MSI-X interrupts,
1316 	 * receive-side processing is executed directly in the interrupt
1317 	 * service routine. Alternatively, we may schedule the taskqueue. */
1318 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
1319 }
1320 
1321 static void
ptnet_vlan_tag_remove(struct mbuf * m)1322 ptnet_vlan_tag_remove(struct mbuf *m)
1323 {
1324 	struct ether_vlan_header *evh;
1325 
1326 	evh = mtod(m, struct ether_vlan_header *);
1327 	m->m_pkthdr.ether_vtag = ntohs(evh->evl_tag);
1328 	m->m_flags |= M_VLANTAG;
1329 
1330 	/* Strip the 802.1Q header. */
1331 	bcopy((char *) evh, (char *) evh + ETHER_VLAN_ENCAP_LEN,
1332 	    ETHER_HDR_LEN - ETHER_TYPE_LEN);
1333 	m_adj(m, ETHER_VLAN_ENCAP_LEN);
1334 }
1335 
1336 static void
ptnet_ring_update(struct ptnet_queue * pq,struct netmap_kring * kring,unsigned int head,unsigned int sync_flags)1337 ptnet_ring_update(struct ptnet_queue *pq, struct netmap_kring *kring,
1338 		  unsigned int head, unsigned int sync_flags)
1339 {
1340 	struct netmap_ring *ring = kring->ring;
1341 	struct nm_csb_atok *atok = pq->atok;
1342 	struct nm_csb_ktoa *ktoa = pq->ktoa;
1343 
1344 	/* Some packets have been pushed to the netmap ring. We have
1345 	 * to tell the host to process the new packets, updating cur
1346 	 * and head in the CSB. */
1347 	ring->head = ring->cur = head;
1348 
1349 	/* Mimic nm_txsync_prologue/nm_rxsync_prologue. */
1350 	kring->rcur = kring->rhead = head;
1351 
1352 	nm_sync_kloop_appl_write(atok, kring->rcur, kring->rhead);
1353 
1354 	/* Kick the host if needed. */
1355 	if (NM_ACCESS_ONCE(ktoa->kern_need_kick)) {
1356 		atok->sync_flags = sync_flags;
1357 		ptnet_kick(pq);
1358 	}
1359 }
1360 
1361 #define PTNET_TX_NOSPACE(_h, _k, _min)	\
1362 	((((_h) < (_k)->rtail) ? 0 : (_k)->nkr_num_slots) + \
1363 		(_k)->rtail - (_h)) < (_min)
1364 
1365 /* This function may be called by the network stack, or by
1366  * by the taskqueue thread. */
1367 static int
ptnet_drain_transmit_queue(struct ptnet_queue * pq,unsigned int budget,bool may_resched)1368 ptnet_drain_transmit_queue(struct ptnet_queue *pq, unsigned int budget,
1369 			   bool may_resched)
1370 {
1371 	struct ptnet_softc *sc = pq->sc;
1372 	bool have_vnet_hdr = sc->vnet_hdr_len;
1373 	struct netmap_adapter *na = &sc->ptna->dr.up;
1374 	if_t ifp = sc->ifp;
1375 	unsigned int batch_count = 0;
1376 	struct nm_csb_atok *atok;
1377 	struct nm_csb_ktoa *ktoa;
1378 	struct netmap_kring *kring;
1379 	struct netmap_ring *ring;
1380 	struct netmap_slot *slot;
1381 	unsigned int count = 0;
1382 	unsigned int minspace;
1383 	unsigned int head;
1384 	unsigned int lim;
1385 	struct mbuf *mhead;
1386 	struct mbuf *mf;
1387 	int nmbuf_bytes;
1388 	uint8_t *nmbuf;
1389 
1390 	if (!PTNET_Q_TRYLOCK(pq)) {
1391 		/* We failed to acquire the lock, schedule the taskqueue. */
1392 		nm_prlim(1, "Deferring TX work");
1393 		if (may_resched) {
1394 			taskqueue_enqueue(pq->taskq, &pq->task);
1395 		}
1396 
1397 		return 0;
1398 	}
1399 
1400 	if (unlikely(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))) {
1401 		PTNET_Q_UNLOCK(pq);
1402 		nm_prlim(1, "Interface is down");
1403 		return ENETDOWN;
1404 	}
1405 
1406 	atok = pq->atok;
1407 	ktoa = pq->ktoa;
1408 	kring = na->tx_rings[pq->kring_id];
1409 	ring = kring->ring;
1410 	lim = kring->nkr_num_slots - 1;
1411 	head = ring->head;
1412 	minspace = sc->min_tx_space;
1413 
1414 	while (count < budget) {
1415 		if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1416 			/* We ran out of slot, let's see if the host has
1417 			 * freed up some, by reading hwcur and hwtail from
1418 			 * the CSB. */
1419 			ptnet_sync_tail(ktoa, kring);
1420 
1421 			if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1422 				/* Still no slots available. Reactivate the
1423 				 * interrupts so that we can be notified
1424 				 * when some free slots are made available by
1425 				 * the host. */
1426 				atok->appl_need_kick = 1;
1427 
1428 				/* Double check. We need a full barrier to
1429 				 * prevent the store to atok->appl_need_kick
1430 				 * to be reordered with the load from
1431 				 * ktoa->hwcur and ktoa->hwtail (store-load
1432 				 * barrier). */
1433 				nm_stld_barrier();
1434 				ptnet_sync_tail(ktoa, kring);
1435 				if (likely(PTNET_TX_NOSPACE(head, kring,
1436 							    minspace))) {
1437 					break;
1438 				}
1439 
1440 				nm_prlim(1, "Found more slots by doublecheck");
1441 				/* More slots were freed before reactivating
1442 				 * the interrupts. */
1443 				atok->appl_need_kick = 0;
1444 			}
1445 		}
1446 
1447 		mhead = drbr_peek(ifp, pq->bufring);
1448 		if (!mhead) {
1449 			break;
1450 		}
1451 
1452 		/* Initialize transmission state variables. */
1453 		slot = ring->slot + head;
1454 		nmbuf = NMB(na, slot);
1455 		nmbuf_bytes = 0;
1456 
1457 		/* If needed, prepare the virtio-net header at the beginning
1458 		 * of the first slot. */
1459 		if (have_vnet_hdr) {
1460 			struct virtio_net_hdr *vh =
1461 					(struct virtio_net_hdr *)nmbuf;
1462 
1463 			/* For performance, we could replace this memset() with
1464 			 * two 8-bytes-wide writes. */
1465 			memset(nmbuf, 0, PTNET_HDR_SIZE);
1466 			if (mhead->m_pkthdr.csum_flags & PTNET_ALL_OFFLOAD) {
1467 				mhead = virtio_net_tx_offload(ifp, mhead, false,
1468 							 vh);
1469 				if (unlikely(!mhead)) {
1470 					/* Packet dropped because errors
1471 					 * occurred while preparing the vnet
1472 					 * header. Let's go ahead with the next
1473 					 * packet. */
1474 					pq->stats.errors ++;
1475 					drbr_advance(ifp, pq->bufring);
1476 					continue;
1477 				}
1478 			}
1479 			nm_prdis(1, "%s: [csum_flags %lX] vnet hdr: flags %x "
1480 			      "csum_start %u csum_ofs %u hdr_len = %u "
1481 			      "gso_size %u gso_type %x", __func__,
1482 			      mhead->m_pkthdr.csum_flags, vh->flags,
1483 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
1484 			      vh->gso_size, vh->gso_type);
1485 
1486 			nmbuf += PTNET_HDR_SIZE;
1487 			nmbuf_bytes += PTNET_HDR_SIZE;
1488 		}
1489 
1490 		for (mf = mhead; mf; mf = mf->m_next) {
1491 			uint8_t *mdata = mf->m_data;
1492 			int mlen = mf->m_len;
1493 
1494 			for (;;) {
1495 				int copy = NETMAP_BUF_SIZE(na) - nmbuf_bytes;
1496 
1497 				if (mlen < copy) {
1498 					copy = mlen;
1499 				}
1500 				memcpy(nmbuf, mdata, copy);
1501 
1502 				mdata += copy;
1503 				mlen -= copy;
1504 				nmbuf += copy;
1505 				nmbuf_bytes += copy;
1506 
1507 				if (!mlen) {
1508 					break;
1509 				}
1510 
1511 				slot->len = nmbuf_bytes;
1512 				slot->flags = NS_MOREFRAG;
1513 
1514 				head = nm_next(head, lim);
1515 				KASSERT(head != ring->tail,
1516 					("Unexpectedly run out of TX space"));
1517 				slot = ring->slot + head;
1518 				nmbuf = NMB(na, slot);
1519 				nmbuf_bytes = 0;
1520 			}
1521 		}
1522 
1523 		/* Complete last slot and update head. */
1524 		slot->len = nmbuf_bytes;
1525 		slot->flags = 0;
1526 		head = nm_next(head, lim);
1527 
1528 		/* Consume the packet just processed. */
1529 		drbr_advance(ifp, pq->bufring);
1530 
1531 		/* Copy the packet to listeners. */
1532 		ETHER_BPF_MTAP(ifp, mhead);
1533 
1534 		pq->stats.packets ++;
1535 		pq->stats.bytes += mhead->m_pkthdr.len;
1536 		if (mhead->m_flags & M_MCAST) {
1537 			pq->stats.mcasts ++;
1538 		}
1539 
1540 		m_freem(mhead);
1541 
1542 		count ++;
1543 		if (++batch_count == PTNET_TX_BATCH) {
1544 			ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1545 			batch_count = 0;
1546 		}
1547 	}
1548 
1549 	if (batch_count) {
1550 		ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1551 	}
1552 
1553 	if (count >= budget && may_resched) {
1554 		DBG(nm_prlim(1, "out of budget: resched, %d mbufs pending\n",
1555 					drbr_inuse(ifp, pq->bufring)));
1556 		taskqueue_enqueue(pq->taskq, &pq->task);
1557 	}
1558 
1559 	PTNET_Q_UNLOCK(pq);
1560 
1561 	return count;
1562 }
1563 
1564 static int
ptnet_transmit(if_t ifp,struct mbuf * m)1565 ptnet_transmit(if_t ifp, struct mbuf *m)
1566 {
1567 	struct ptnet_softc *sc = if_getsoftc(ifp);
1568 	struct ptnet_queue *pq;
1569 	unsigned int queue_idx;
1570 	int err;
1571 
1572 	DBG(device_printf(sc->dev, "transmit %p\n", m));
1573 
1574 	/* Insert 802.1Q header if needed. */
1575 	if (m->m_flags & M_VLANTAG) {
1576 		m = ether_vlanencap(m, m->m_pkthdr.ether_vtag);
1577 		if (m == NULL) {
1578 			return ENOBUFS;
1579 		}
1580 		m->m_flags &= ~M_VLANTAG;
1581 	}
1582 
1583 	/* Get the flow-id if available. */
1584 	queue_idx = (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) ?
1585 		    m->m_pkthdr.flowid : curcpu;
1586 
1587 	if (unlikely(queue_idx >= sc->num_tx_rings)) {
1588 		queue_idx %= sc->num_tx_rings;
1589 	}
1590 
1591 	pq = sc->queues + queue_idx;
1592 
1593 	err = drbr_enqueue(ifp, pq->bufring, m);
1594 	if (err) {
1595 		/* ENOBUFS when the bufring is full */
1596 		nm_prlim(1, "%s: drbr_enqueue() failed %d\n",
1597 			__func__, err);
1598 		pq->stats.errors ++;
1599 		return err;
1600 	}
1601 
1602 	if (if_getcapenable(ifp) & IFCAP_POLLING) {
1603 		/* If polling is on, the transmit queues will be
1604 		 * drained by the poller. */
1605 		return 0;
1606 	}
1607 
1608 	err = ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
1609 
1610 	return (err < 0) ? err : 0;
1611 }
1612 
1613 static unsigned int
ptnet_rx_discard(struct netmap_kring * kring,unsigned int head)1614 ptnet_rx_discard(struct netmap_kring *kring, unsigned int head)
1615 {
1616 	struct netmap_ring *ring = kring->ring;
1617 	struct netmap_slot *slot = ring->slot + head;
1618 
1619 	for (;;) {
1620 		head = nm_next(head, kring->nkr_num_slots - 1);
1621 		if (!(slot->flags & NS_MOREFRAG) || head == ring->tail) {
1622 			break;
1623 		}
1624 		slot = ring->slot + head;
1625 	}
1626 
1627 	return head;
1628 }
1629 
1630 static inline struct mbuf *
ptnet_rx_slot(struct mbuf * mtail,uint8_t * nmbuf,unsigned int nmbuf_len)1631 ptnet_rx_slot(struct mbuf *mtail, uint8_t *nmbuf, unsigned int nmbuf_len)
1632 {
1633 	uint8_t *mdata = mtod(mtail, uint8_t *) + mtail->m_len;
1634 
1635 	do {
1636 		unsigned int copy;
1637 
1638 		if (mtail->m_len == MCLBYTES) {
1639 			struct mbuf *mf;
1640 
1641 			mf = m_getcl(M_NOWAIT, MT_DATA, 0);
1642 			if (unlikely(!mf)) {
1643 				return NULL;
1644 			}
1645 
1646 			mtail->m_next = mf;
1647 			mtail = mf;
1648 			mdata = mtod(mtail, uint8_t *);
1649 			mtail->m_len = 0;
1650 		}
1651 
1652 		copy = MCLBYTES - mtail->m_len;
1653 		if (nmbuf_len < copy) {
1654 			copy = nmbuf_len;
1655 		}
1656 
1657 		memcpy(mdata, nmbuf, copy);
1658 
1659 		nmbuf += copy;
1660 		nmbuf_len -= copy;
1661 		mdata += copy;
1662 		mtail->m_len += copy;
1663 	} while (nmbuf_len);
1664 
1665 	return mtail;
1666 }
1667 
1668 static int
ptnet_rx_eof(struct ptnet_queue * pq,unsigned int budget,bool may_resched)1669 ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget, bool may_resched)
1670 {
1671 	struct ptnet_softc *sc = pq->sc;
1672 	bool have_vnet_hdr = sc->vnet_hdr_len;
1673 	struct nm_csb_atok *atok = pq->atok;
1674 	struct nm_csb_ktoa *ktoa = pq->ktoa;
1675 	struct netmap_adapter *na = &sc->ptna->dr.up;
1676 	struct netmap_kring *kring = na->rx_rings[pq->kring_id];
1677 	struct netmap_ring *ring = kring->ring;
1678 	unsigned int const lim = kring->nkr_num_slots - 1;
1679 	unsigned int batch_count = 0;
1680 	if_t ifp = sc->ifp;
1681 	unsigned int count = 0;
1682 	uint32_t head;
1683 
1684 	PTNET_Q_LOCK(pq);
1685 
1686 	if (unlikely(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))) {
1687 		goto unlock;
1688 	}
1689 
1690 	kring->nr_kflags &= ~NKR_PENDINTR;
1691 
1692 	head = ring->head;
1693 	while (count < budget) {
1694 		uint32_t prev_head = head;
1695 		struct mbuf *mhead, *mtail;
1696 		struct virtio_net_hdr *vh;
1697 		struct netmap_slot *slot;
1698 		unsigned int nmbuf_len;
1699 		uint8_t *nmbuf;
1700 		int deliver = 1; /* the mbuf to the network stack. */
1701 host_sync:
1702 		if (head == ring->tail) {
1703 			/* We ran out of slot, let's see if the host has
1704 			 * added some, by reading hwcur and hwtail from
1705 			 * the CSB. */
1706 			ptnet_sync_tail(ktoa, kring);
1707 
1708 			if (head == ring->tail) {
1709 				/* Still no slots available. Reactivate
1710 				 * interrupts as they were disabled by the
1711 				 * host thread right before issuing the
1712 				 * last interrupt. */
1713 				atok->appl_need_kick = 1;
1714 
1715 				/* Double check for more completed RX slots.
1716 				 * We need a full barrier to prevent the store
1717 				 * to atok->appl_need_kick to be reordered with
1718 				 * the load from ktoa->hwcur and ktoa->hwtail
1719 				 * (store-load barrier). */
1720 				nm_stld_barrier();
1721 				ptnet_sync_tail(ktoa, kring);
1722 				if (likely(head == ring->tail)) {
1723 					break;
1724 				}
1725 				atok->appl_need_kick = 0;
1726 			}
1727 		}
1728 
1729 		/* Initialize ring state variables, possibly grabbing the
1730 		 * virtio-net header. */
1731 		slot = ring->slot + head;
1732 		nmbuf = NMB(na, slot);
1733 		nmbuf_len = slot->len;
1734 
1735 		vh = (struct virtio_net_hdr *)nmbuf;
1736 		if (have_vnet_hdr) {
1737 			if (unlikely(nmbuf_len < PTNET_HDR_SIZE)) {
1738 				/* There is no good reason why host should
1739 				 * put the header in multiple netmap slots.
1740 				 * If this is the case, discard. */
1741 				nm_prlim(1, "Fragmented vnet-hdr: dropping");
1742 				head = ptnet_rx_discard(kring, head);
1743 				pq->stats.iqdrops ++;
1744 				deliver = 0;
1745 				goto skip;
1746 			}
1747 			nm_prdis(1, "%s: vnet hdr: flags %x csum_start %u "
1748 			      "csum_ofs %u hdr_len = %u gso_size %u "
1749 			      "gso_type %x", __func__, vh->flags,
1750 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
1751 			      vh->gso_size, vh->gso_type);
1752 			nmbuf += PTNET_HDR_SIZE;
1753 			nmbuf_len -= PTNET_HDR_SIZE;
1754 		}
1755 
1756 		/* Allocate the head of a new mbuf chain.
1757 		 * We use m_getcl() to allocate an mbuf with standard cluster
1758 		 * size (MCLBYTES). In the future we could use m_getjcl()
1759 		 * to choose different sizes. */
1760 		mhead = mtail = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1761 		if (unlikely(mhead == NULL)) {
1762 			device_printf(sc->dev, "%s: failed to allocate mbuf "
1763 				      "head\n", __func__);
1764 			pq->stats.errors ++;
1765 			break;
1766 		}
1767 
1768 		/* Initialize the mbuf state variables. */
1769 		mhead->m_pkthdr.len = nmbuf_len;
1770 		mtail->m_len = 0;
1771 
1772 		/* Scan all the netmap slots containing the current packet. */
1773 		for (;;) {
1774 			DBG(device_printf(sc->dev, "%s: h %u t %u rcv frag "
1775 					  "len %u, flags %u\n", __func__,
1776 					  head, ring->tail, slot->len,
1777 					  slot->flags));
1778 
1779 			mtail = ptnet_rx_slot(mtail, nmbuf, nmbuf_len);
1780 			if (unlikely(!mtail)) {
1781 				/* Ouch. We ran out of memory while processing
1782 				 * a packet. We have to restore the previous
1783 				 * head position, free the mbuf chain, and
1784 				 * schedule the taskqueue to give the packet
1785 				 * another chance. */
1786 				device_printf(sc->dev, "%s: failed to allocate"
1787 					" mbuf frag, reset head %u --> %u\n",
1788 					__func__, head, prev_head);
1789 				head = prev_head;
1790 				m_freem(mhead);
1791 				pq->stats.errors ++;
1792 				if (may_resched) {
1793 					taskqueue_enqueue(pq->taskq,
1794 							  &pq->task);
1795 				}
1796 				goto escape;
1797 			}
1798 
1799 			/* We have to increment head irrespective of the
1800 			 * NS_MOREFRAG being set or not. */
1801 			head = nm_next(head, lim);
1802 
1803 			if (!(slot->flags & NS_MOREFRAG)) {
1804 				break;
1805 			}
1806 
1807 			if (unlikely(head == ring->tail)) {
1808 				/* The very last slot prepared by the host has
1809 				 * the NS_MOREFRAG set. Drop it and continue
1810 				 * the outer cycle (to do the double-check). */
1811 				nm_prlim(1, "Incomplete packet: dropping");
1812 				m_freem(mhead);
1813 				pq->stats.iqdrops ++;
1814 				goto host_sync;
1815 			}
1816 
1817 			slot = ring->slot + head;
1818 			nmbuf = NMB(na, slot);
1819 			nmbuf_len = slot->len;
1820 			mhead->m_pkthdr.len += nmbuf_len;
1821 		}
1822 
1823 		mhead->m_pkthdr.rcvif = ifp;
1824 		mhead->m_pkthdr.csum_flags = 0;
1825 
1826 		/* Store the queue idx in the packet header. */
1827 		mhead->m_pkthdr.flowid = pq->kring_id;
1828 		M_HASHTYPE_SET(mhead, M_HASHTYPE_OPAQUE);
1829 
1830 		if (if_getcapenable(ifp) & IFCAP_VLAN_HWTAGGING) {
1831 			struct ether_header *eh;
1832 
1833 			eh = mtod(mhead, struct ether_header *);
1834 			if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
1835 				ptnet_vlan_tag_remove(mhead);
1836 				/*
1837 				 * With the 802.1Q header removed, update the
1838 				 * checksum starting location accordingly.
1839 				 */
1840 				if (vh->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
1841 					vh->csum_start -= ETHER_VLAN_ENCAP_LEN;
1842 			}
1843 		}
1844 
1845 		if (unlikely(have_vnet_hdr && virtio_net_rx_csum(mhead, vh))) {
1846 			m_freem(mhead);
1847 			nm_prlim(1, "Csum offload error: dropping");
1848 			pq->stats.iqdrops ++;
1849 			deliver = 0;
1850 		}
1851 
1852 skip:
1853 		count ++;
1854 		if (++batch_count >= PTNET_RX_BATCH) {
1855 			/* Some packets have been (or will be) pushed to the network
1856 			 * stack. We need to update the CSB to tell the host about
1857 			 * the new ring->cur and ring->head (RX buffer refill). */
1858 			ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
1859 			batch_count = 0;
1860 		}
1861 
1862 		if (likely(deliver))  {
1863 			pq->stats.packets ++;
1864 			pq->stats.bytes += mhead->m_pkthdr.len;
1865 
1866 			PTNET_Q_UNLOCK(pq);
1867 			if_input(ifp, mhead);
1868 			PTNET_Q_LOCK(pq);
1869 			/* The ring->head index (and related indices) are
1870 			 * updated under pq lock by ptnet_ring_update().
1871 			 * Since we dropped the lock to call if_input(), we
1872 			 * must reload ring->head and restart processing the
1873 			 * ring from there. */
1874 			head = ring->head;
1875 
1876 			if (unlikely(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))) {
1877 				/* The interface has gone down while we didn't
1878 				 * have the lock. Stop any processing and exit. */
1879 				goto unlock;
1880 			}
1881 		}
1882 	}
1883 escape:
1884 	if (batch_count) {
1885 		ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
1886 
1887 	}
1888 
1889 	if (count >= budget && may_resched) {
1890 		/* If we ran out of budget or the double-check found new
1891 		 * slots to process, schedule the taskqueue. */
1892 		DBG(nm_prlim(1, "out of budget: resched h %u t %u\n",
1893 					head, ring->tail));
1894 		taskqueue_enqueue(pq->taskq, &pq->task);
1895 	}
1896 unlock:
1897 	PTNET_Q_UNLOCK(pq);
1898 
1899 	return count;
1900 }
1901 
1902 static void
ptnet_rx_task(void * context,int pending)1903 ptnet_rx_task(void *context, int pending)
1904 {
1905 	struct ptnet_queue *pq = context;
1906 
1907 	DBG(nm_prlim(1, "%s: pq #%u\n", __func__, pq->kring_id));
1908 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
1909 }
1910 
1911 static void
ptnet_tx_task(void * context,int pending)1912 ptnet_tx_task(void *context, int pending)
1913 {
1914 	struct ptnet_queue *pq = context;
1915 
1916 	DBG(nm_prlim(1, "%s: pq #%u\n", __func__, pq->kring_id));
1917 	ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
1918 }
1919 
1920 #ifdef DEVICE_POLLING
1921 /* We don't need to handle differently POLL_AND_CHECK_STATUS and
1922  * POLL_ONLY, since we don't have an Interrupt Status Register. */
1923 static int
ptnet_poll(if_t ifp,enum poll_cmd cmd,int budget)1924 ptnet_poll(if_t ifp, enum poll_cmd cmd, int budget)
1925 {
1926 	struct ptnet_softc *sc = if_getsoftc(ifp);
1927 	unsigned int queue_budget;
1928 	unsigned int count = 0;
1929 	bool borrow = false;
1930 	int i;
1931 
1932 	KASSERT(sc->num_rings > 0, ("Found no queues in while polling ptnet"));
1933 	queue_budget = MAX(budget / sc->num_rings, 1);
1934 	nm_prlim(1, "Per-queue budget is %d", queue_budget);
1935 
1936 	while (budget) {
1937 		unsigned int rcnt = 0;
1938 
1939 		for (i = 0; i < sc->num_rings; i++) {
1940 			struct ptnet_queue *pq = sc->queues + i;
1941 
1942 			if (borrow) {
1943 				queue_budget = MIN(queue_budget, budget);
1944 				if (queue_budget == 0) {
1945 					break;
1946 				}
1947 			}
1948 
1949 			if (i < sc->num_tx_rings) {
1950 				rcnt += ptnet_drain_transmit_queue(pq,
1951 						   queue_budget, false);
1952 			} else {
1953 				rcnt += ptnet_rx_eof(pq, queue_budget,
1954 						      false);
1955 			}
1956 		}
1957 
1958 		if (!rcnt) {
1959 			/* A scan of the queues gave no result, we can
1960 			 * stop here. */
1961 			break;
1962 		}
1963 
1964 		if (rcnt > budget) {
1965 			/* This may happen when initial budget < sc->num_rings,
1966 			 * since one packet budget is given to each queue
1967 			 * anyway. Just pretend we didn't eat "so much". */
1968 			rcnt = budget;
1969 		}
1970 		count += rcnt;
1971 		budget -= rcnt;
1972 		borrow = true;
1973 	}
1974 
1975 
1976 	return count;
1977 }
1978 #endif /* DEVICE_POLLING */
1979 #endif /* WITH_PTNETMAP */
1980