xref: /freebsd-13-stable/sys/dev/netmap/if_ptnet.c (revision f500e5c6c99bd4520daa4524113462e3cf68f032)
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 
31 #include <sys/types.h>
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/kernel.h>
35 #include <sys/sockio.h>
36 #include <sys/mbuf.h>
37 #include <sys/malloc.h>
38 #include <sys/module.h>
39 #include <sys/socket.h>
40 #include <sys/sysctl.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/taskqueue.h>
44 #include <sys/smp.h>
45 #include <sys/time.h>
46 #include <machine/smp.h>
47 
48 #include <vm/uma.h>
49 #include <vm/vm.h>
50 #include <vm/pmap.h>
51 
52 #include <net/ethernet.h>
53 #include <net/if.h>
54 #include <net/if_var.h>
55 #include <net/if_arp.h>
56 #include <net/if_dl.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/if_vlan_var.h>
60 #include <net/bpf.h>
61 
62 #include <netinet/in_systm.h>
63 #include <netinet/in.h>
64 #include <netinet/ip.h>
65 #include <netinet/ip6.h>
66 #include <netinet6/ip6_var.h>
67 #include <netinet/udp.h>
68 #include <netinet/tcp.h>
69 
70 #include <machine/bus.h>
71 #include <machine/resource.h>
72 #include <sys/bus.h>
73 #include <sys/rman.h>
74 
75 #include <dev/pci/pcivar.h>
76 #include <dev/pci/pcireg.h>
77 
78 #include "opt_inet.h"
79 #include "opt_inet6.h"
80 
81 #include <sys/selinfo.h>
82 #include <net/netmap.h>
83 #include <dev/netmap/netmap_kern.h>
84 #include <net/netmap_virt.h>
85 #include <dev/netmap/netmap_mem2.h>
86 #include <dev/virtio/network/virtio_net.h>
87 
88 #ifdef WITH_PTNETMAP
89 
90 #ifndef INET
91 #error "INET not defined, cannot support offloadings"
92 #endif
93 
94 static uint64_t	ptnet_get_counter(if_t, ift_counter);
95 
96 //#define PTNETMAP_STATS
97 //#define DEBUG
98 #ifdef DEBUG
99 #define DBG(x) x
100 #else   /* !DEBUG */
101 #define DBG(x)
102 #endif  /* !DEBUG */
103 
104 extern int ptnet_vnet_hdr; /* Tunable parameter */
105 
106 struct ptnet_softc;
107 
108 struct ptnet_queue_stats {
109 	uint64_t	packets; /* if_[io]packets */
110 	uint64_t	bytes;	 /* if_[io]bytes */
111 	uint64_t	errors;	 /* if_[io]errors */
112 	uint64_t	iqdrops; /* if_iqdrops */
113 	uint64_t	mcasts;  /* if_[io]mcasts */
114 #ifdef PTNETMAP_STATS
115 	uint64_t	intrs;
116 	uint64_t	kicks;
117 #endif /* PTNETMAP_STATS */
118 };
119 
120 struct ptnet_queue {
121 	struct ptnet_softc		*sc;
122 	struct				resource *irq;
123 	void				*cookie;
124 	int				kring_id;
125 	struct nm_csb_atok		*atok;
126 	struct nm_csb_ktoa		*ktoa;
127 	unsigned int			kick;
128 	struct mtx			lock;
129 	struct buf_ring			*bufring; /* for TX queues */
130 	struct ptnet_queue_stats	stats;
131 #ifdef PTNETMAP_STATS
132 	struct ptnet_queue_stats	last_stats;
133 #endif /* PTNETMAP_STATS */
134 	struct taskqueue		*taskq;
135 	struct task			task;
136 	char				lock_name[16];
137 };
138 
139 #define PTNET_Q_LOCK(_pq)	mtx_lock(&(_pq)->lock)
140 #define PTNET_Q_TRYLOCK(_pq)	mtx_trylock(&(_pq)->lock)
141 #define PTNET_Q_UNLOCK(_pq)	mtx_unlock(&(_pq)->lock)
142 
143 struct ptnet_softc {
144 	device_t		dev;
145 	if_t			ifp;
146 	struct ifmedia		media;
147 	struct mtx		lock;
148 	char			lock_name[16];
149 	char			hwaddr[ETHER_ADDR_LEN];
150 
151 	/* Mirror of PTFEAT register. */
152 	uint32_t		ptfeatures;
153 	unsigned int		vnet_hdr_len;
154 
155 	/* PCI BARs support. */
156 	struct resource		*iomem;
157 	struct resource		*msix_mem;
158 
159 	unsigned int		num_rings;
160 	unsigned int		num_tx_rings;
161 	struct ptnet_queue	*queues;
162 	struct ptnet_queue	*rxqueues;
163 	struct nm_csb_atok	*csb_gh;
164 	struct nm_csb_ktoa	*csb_hg;
165 
166 	unsigned int		min_tx_space;
167 
168 	struct netmap_pt_guest_adapter *ptna;
169 
170 	struct callout		tick;
171 #ifdef PTNETMAP_STATS
172 	struct timeval		last_ts;
173 #endif /* PTNETMAP_STATS */
174 };
175 
176 #define PTNET_CORE_LOCK(_sc)	mtx_lock(&(_sc)->lock)
177 #define PTNET_CORE_UNLOCK(_sc)	mtx_unlock(&(_sc)->lock)
178 
179 static int	ptnet_probe(device_t);
180 static int	ptnet_attach(device_t);
181 static int	ptnet_detach(device_t);
182 static int	ptnet_suspend(device_t);
183 static int	ptnet_resume(device_t);
184 static int	ptnet_shutdown(device_t);
185 
186 static void	ptnet_init(void *opaque);
187 static int	ptnet_ioctl(if_t ifp, u_long cmd, caddr_t data);
188 static int	ptnet_init_locked(struct ptnet_softc *sc);
189 static int	ptnet_stop(struct ptnet_softc *sc);
190 static int	ptnet_transmit(if_t ifp, struct mbuf *m);
191 static int	ptnet_drain_transmit_queue(struct ptnet_queue *pq,
192 					   unsigned int budget,
193 					   bool may_resched);
194 static void	ptnet_qflush(if_t ifp);
195 static void	ptnet_tx_task(void *context, int pending);
196 
197 static int	ptnet_media_change(if_t ifp);
198 static void	ptnet_media_status(if_t ifp, struct ifmediareq *ifmr);
199 #ifdef PTNETMAP_STATS
200 static void	ptnet_tick(void *opaque);
201 #endif
202 
203 static int	ptnet_irqs_init(struct ptnet_softc *sc);
204 static void	ptnet_irqs_fini(struct ptnet_softc *sc);
205 
206 static uint32_t ptnet_nm_ptctl(struct ptnet_softc *sc, uint32_t cmd);
207 static int      ptnet_nm_config(struct netmap_adapter *na,
208 				struct nm_config_info *info);
209 static void	ptnet_update_vnet_hdr(struct ptnet_softc *sc);
210 static int	ptnet_nm_register(struct netmap_adapter *na, int onoff);
211 static int	ptnet_nm_txsync(struct netmap_kring *kring, int flags);
212 static int	ptnet_nm_rxsync(struct netmap_kring *kring, int flags);
213 static void	ptnet_nm_intr(struct netmap_adapter *na, int onoff);
214 
215 static void	ptnet_tx_intr(void *opaque);
216 static void	ptnet_rx_intr(void *opaque);
217 
218 static unsigned	ptnet_rx_discard(struct netmap_kring *kring,
219 				 unsigned int head);
220 static int	ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget,
221 			     bool may_resched);
222 static void	ptnet_rx_task(void *context, int pending);
223 
224 #ifdef DEVICE_POLLING
225 static poll_handler_t ptnet_poll;
226 #endif
227 
228 static device_method_t ptnet_methods[] = {
229 	DEVMETHOD(device_probe,			ptnet_probe),
230 	DEVMETHOD(device_attach,		ptnet_attach),
231 	DEVMETHOD(device_detach,		ptnet_detach),
232 	DEVMETHOD(device_suspend,		ptnet_suspend),
233 	DEVMETHOD(device_resume,		ptnet_resume),
234 	DEVMETHOD(device_shutdown,		ptnet_shutdown),
235 	DEVMETHOD_END
236 };
237 
238 static driver_t ptnet_driver = {
239 	"ptnet",
240 	ptnet_methods,
241 	sizeof(struct ptnet_softc)
242 };
243 
244 /* We use (SI_ORDER_MIDDLE+2) here, see DEV_MODULE_ORDERED() invocation. */
245 static devclass_t ptnet_devclass;
246 DRIVER_MODULE_ORDERED(ptnet, pci, ptnet_driver, ptnet_devclass,
247 		      NULL, NULL, SI_ORDER_MIDDLE + 2);
248 
249 static int
ptnet_probe(device_t dev)250 ptnet_probe(device_t dev)
251 {
252 	if (pci_get_vendor(dev) != PTNETMAP_PCI_VENDOR_ID ||
253 		pci_get_device(dev) != PTNETMAP_PCI_NETIF_ID) {
254 		return (ENXIO);
255 	}
256 
257 	device_set_desc(dev, "ptnet network adapter");
258 
259 	return (BUS_PROBE_DEFAULT);
260 }
261 
ptnet_kick(struct ptnet_queue * pq)262 static inline void ptnet_kick(struct ptnet_queue *pq)
263 {
264 #ifdef PTNETMAP_STATS
265 	pq->stats.kicks ++;
266 #endif /* PTNETMAP_STATS */
267 	bus_write_4(pq->sc->iomem, pq->kick, 0);
268 }
269 
270 #define PTNET_BUF_RING_SIZE	4096
271 #define PTNET_RX_BUDGET		512
272 #define PTNET_RX_BATCH		1
273 #define PTNET_TX_BUDGET		512
274 #define PTNET_TX_BATCH		64
275 #define PTNET_HDR_SIZE		sizeof(struct virtio_net_hdr_mrg_rxbuf)
276 #define PTNET_MAX_PKT_SIZE	65536
277 
278 #define PTNET_CSUM_OFFLOAD	(CSUM_TCP | CSUM_UDP)
279 #define PTNET_CSUM_OFFLOAD_IPV6	(CSUM_TCP_IPV6 | CSUM_UDP_IPV6)
280 #define PTNET_ALL_OFFLOAD	(CSUM_TSO | PTNET_CSUM_OFFLOAD |\
281 				 PTNET_CSUM_OFFLOAD_IPV6)
282 
283 static int
ptnet_attach(device_t dev)284 ptnet_attach(device_t dev)
285 {
286 	uint32_t ptfeatures = 0;
287 	unsigned int num_rx_rings, num_tx_rings;
288 	struct netmap_adapter na_arg;
289 	unsigned int nifp_offset;
290 	struct ptnet_softc *sc;
291 	if_t ifp;
292 	uint32_t macreg;
293 	int err, rid;
294 	int i;
295 
296 	sc = device_get_softc(dev);
297 	sc->dev = dev;
298 
299 	/* Setup PCI resources. */
300 	pci_enable_busmaster(dev);
301 
302 	rid = PCIR_BAR(PTNETMAP_IO_PCI_BAR);
303 	sc->iomem = bus_alloc_resource_any(dev, SYS_RES_IOPORT, &rid,
304 					   RF_ACTIVE);
305 	if (sc->iomem == NULL) {
306 		device_printf(dev, "Failed to map I/O BAR\n");
307 		return (ENXIO);
308 	}
309 
310 	/* Negotiate features with the hypervisor. */
311 	if (ptnet_vnet_hdr) {
312 		ptfeatures |= PTNETMAP_F_VNET_HDR;
313 	}
314 	bus_write_4(sc->iomem, PTNET_IO_PTFEAT, ptfeatures); /* wanted */
315 	ptfeatures = bus_read_4(sc->iomem, PTNET_IO_PTFEAT); /* acked */
316 	sc->ptfeatures = ptfeatures;
317 
318 	num_tx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_RINGS);
319 	num_rx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_RINGS);
320 	sc->num_rings = num_tx_rings + num_rx_rings;
321 	sc->num_tx_rings = num_tx_rings;
322 
323 	if (sc->num_rings * sizeof(struct nm_csb_atok) > PAGE_SIZE) {
324 		device_printf(dev, "CSB cannot handle that many rings (%u)\n",
325 				sc->num_rings);
326 		err = ENOMEM;
327 		goto err_path;
328 	}
329 
330 	/* Allocate CSB and carry out CSB allocation protocol. */
331 	sc->csb_gh = contigmalloc(2*PAGE_SIZE, M_DEVBUF, M_NOWAIT | M_ZERO,
332 				  (size_t)0, -1UL, PAGE_SIZE, 0);
333 	if (sc->csb_gh == NULL) {
334 		device_printf(dev, "Failed to allocate CSB\n");
335 		err = ENOMEM;
336 		goto err_path;
337 	}
338 	sc->csb_hg = (struct nm_csb_ktoa *)(((char *)sc->csb_gh) + PAGE_SIZE);
339 
340 	{
341 		/*
342 		 * We use uint64_t rather than vm_paddr_t since we
343 		 * need 64 bit addresses even on 32 bit platforms.
344 		 */
345 		uint64_t paddr = vtophys(sc->csb_gh);
346 
347 		/* CSB allocation protocol: write to BAH first, then
348 		 * to BAL (for both GH and HG sections). */
349 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAH,
350 				(paddr >> 32) & 0xffffffff);
351 		bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAL,
352 				paddr & 0xffffffff);
353 		paddr = vtophys(sc->csb_hg);
354 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAH,
355 				(paddr >> 32) & 0xffffffff);
356 		bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAL,
357 				paddr & 0xffffffff);
358 	}
359 
360 	/* Allocate and initialize per-queue data structures. */
361 	sc->queues = malloc(sizeof(struct ptnet_queue) * sc->num_rings,
362 			    M_DEVBUF, M_NOWAIT | M_ZERO);
363 	if (sc->queues == NULL) {
364 		err = ENOMEM;
365 		goto err_path;
366 	}
367 	sc->rxqueues = sc->queues + num_tx_rings;
368 
369 	for (i = 0; i < sc->num_rings; i++) {
370 		struct ptnet_queue *pq = sc->queues + i;
371 
372 		pq->sc = sc;
373 		pq->kring_id = i;
374 		pq->kick = PTNET_IO_KICK_BASE + 4 * i;
375 		pq->atok = sc->csb_gh + i;
376 		pq->ktoa = sc->csb_hg + i;
377 		snprintf(pq->lock_name, sizeof(pq->lock_name), "%s-%d",
378 			 device_get_nameunit(dev), i);
379 		mtx_init(&pq->lock, pq->lock_name, NULL, MTX_DEF);
380 		if (i >= num_tx_rings) {
381 			/* RX queue: fix kring_id. */
382 			pq->kring_id -= num_tx_rings;
383 		} else {
384 			/* TX queue: allocate buf_ring. */
385 			pq->bufring = buf_ring_alloc(PTNET_BUF_RING_SIZE,
386 						M_DEVBUF, M_NOWAIT, &pq->lock);
387 			if (pq->bufring == NULL) {
388 				err = ENOMEM;
389 				goto err_path;
390 			}
391 		}
392 	}
393 
394 	sc->min_tx_space = 64; /* Safe initial value. */
395 
396 	err = ptnet_irqs_init(sc);
397 	if (err) {
398 		goto err_path;
399 	}
400 
401 	/* Setup Ethernet interface. */
402 	sc->ifp = ifp = if_alloc(IFT_ETHER);
403 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
404 	ifp->if_baudrate = IF_Gbps(10);
405 	ifp->if_softc = sc;
406 	ifp->if_flags = IFF_BROADCAST | IFF_MULTICAST | IFF_SIMPLEX;
407 	ifp->if_init = ptnet_init;
408 	ifp->if_ioctl = ptnet_ioctl;
409 	ifp->if_get_counter = ptnet_get_counter;
410 	ifp->if_transmit = ptnet_transmit;
411 	ifp->if_qflush = ptnet_qflush;
412 
413 	ifmedia_init(&sc->media, IFM_IMASK, ptnet_media_change,
414 		     ptnet_media_status);
415 	ifmedia_add(&sc->media, IFM_ETHER | IFM_10G_T | IFM_FDX, 0, NULL);
416 	ifmedia_set(&sc->media, IFM_ETHER | IFM_10G_T | IFM_FDX);
417 
418 	macreg = bus_read_4(sc->iomem, PTNET_IO_MAC_HI);
419 	sc->hwaddr[0] = (macreg >> 8) & 0xff;
420 	sc->hwaddr[1] = macreg & 0xff;
421 	macreg = bus_read_4(sc->iomem, PTNET_IO_MAC_LO);
422 	sc->hwaddr[2] = (macreg >> 24) & 0xff;
423 	sc->hwaddr[3] = (macreg >> 16) & 0xff;
424 	sc->hwaddr[4] = (macreg >> 8) & 0xff;
425 	sc->hwaddr[5] = macreg & 0xff;
426 
427 	ether_ifattach(ifp, sc->hwaddr);
428 
429 	ifp->if_hdrlen = sizeof(struct ether_vlan_header);
430 	ifp->if_capabilities |= IFCAP_JUMBO_MTU | IFCAP_VLAN_MTU;
431 
432 	if (sc->ptfeatures & PTNETMAP_F_VNET_HDR) {
433 		/* Similarly to what the vtnet driver does, we can emulate
434 		 * VLAN offloadings by inserting and removing the 802.1Q
435 		 * header during transmit and receive. We are then able
436 		 * to do checksum offloading of VLAN frames. */
437 		ifp->if_capabilities |= IFCAP_HWCSUM | IFCAP_HWCSUM_IPV6
438 					| IFCAP_VLAN_HWCSUM
439 					| IFCAP_TSO | IFCAP_LRO
440 					| IFCAP_VLAN_HWTSO
441 					| IFCAP_VLAN_HWTAGGING;
442 	}
443 
444 	ifp->if_capenable = ifp->if_capabilities;
445 #ifdef DEVICE_POLLING
446 	/* Don't enable polling by default. */
447 	ifp->if_capabilities |= IFCAP_POLLING;
448 #endif
449 	snprintf(sc->lock_name, sizeof(sc->lock_name),
450 		 "%s", device_get_nameunit(dev));
451 	mtx_init(&sc->lock, sc->lock_name, "ptnet core lock", MTX_DEF);
452 	callout_init_mtx(&sc->tick, &sc->lock, 0);
453 
454 	/* Prepare a netmap_adapter struct instance to do netmap_attach(). */
455 	nifp_offset = bus_read_4(sc->iomem, PTNET_IO_NIFP_OFS);
456 	memset(&na_arg, 0, sizeof(na_arg));
457 	na_arg.ifp = ifp;
458 	na_arg.num_tx_desc = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_SLOTS);
459 	na_arg.num_rx_desc = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_SLOTS);
460 	na_arg.num_tx_rings = num_tx_rings;
461 	na_arg.num_rx_rings = num_rx_rings;
462 	na_arg.nm_config = ptnet_nm_config;
463 	na_arg.nm_krings_create = ptnet_nm_krings_create;
464 	na_arg.nm_krings_delete = ptnet_nm_krings_delete;
465 	na_arg.nm_dtor = ptnet_nm_dtor;
466 	na_arg.nm_intr = ptnet_nm_intr;
467 	na_arg.nm_register = ptnet_nm_register;
468 	na_arg.nm_txsync = ptnet_nm_txsync;
469 	na_arg.nm_rxsync = ptnet_nm_rxsync;
470 
471 	netmap_pt_guest_attach(&na_arg, nifp_offset,
472                                 bus_read_4(sc->iomem, PTNET_IO_HOSTMEMID));
473 
474 	/* Now a netmap adapter for this ifp has been allocated, and it
475 	 * can be accessed through NA(ifp). We also have to initialize the CSB
476 	 * pointer. */
477 	sc->ptna = (struct netmap_pt_guest_adapter *)NA(ifp);
478 
479 	/* If virtio-net header was negotiated, set the virt_hdr_len field in
480 	 * the netmap adapter, to inform users that this netmap adapter requires
481 	 * the application to deal with the headers. */
482 	ptnet_update_vnet_hdr(sc);
483 
484 	device_printf(dev, "%s() completed\n", __func__);
485 
486 	return (0);
487 
488 err_path:
489 	ptnet_detach(dev);
490 	return err;
491 }
492 
493 /* Stop host sync-kloop if it was running. */
494 static void
ptnet_device_shutdown(struct ptnet_softc * sc)495 ptnet_device_shutdown(struct ptnet_softc *sc)
496 {
497 	ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_DELETE);
498 	bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAH, 0);
499 	bus_write_4(sc->iomem, PTNET_IO_CSB_GH_BAL, 0);
500 	bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAH, 0);
501 	bus_write_4(sc->iomem, PTNET_IO_CSB_HG_BAL, 0);
502 }
503 
504 static int
ptnet_detach(device_t dev)505 ptnet_detach(device_t dev)
506 {
507 	struct ptnet_softc *sc = device_get_softc(dev);
508 	int i;
509 
510 	ptnet_device_shutdown(sc);
511 
512 #ifdef DEVICE_POLLING
513 	if (sc->ifp->if_capenable & IFCAP_POLLING) {
514 		ether_poll_deregister(sc->ifp);
515 	}
516 #endif
517 	callout_drain(&sc->tick);
518 
519 	if (sc->queues) {
520 		/* Drain taskqueues before calling if_detach. */
521 		for (i = 0; i < sc->num_rings; i++) {
522 			struct ptnet_queue *pq = sc->queues + i;
523 
524 			if (pq->taskq) {
525 				taskqueue_drain(pq->taskq, &pq->task);
526 			}
527 		}
528 	}
529 
530 	if (sc->ifp) {
531 		ether_ifdetach(sc->ifp);
532 
533 		/* Uninitialize netmap adapters for this device. */
534 		netmap_detach(sc->ifp);
535 
536 		ifmedia_removeall(&sc->media);
537 		if_free(sc->ifp);
538 		sc->ifp = NULL;
539 	}
540 
541 	ptnet_irqs_fini(sc);
542 
543 	if (sc->csb_gh) {
544 		contigfree(sc->csb_gh, 2*PAGE_SIZE, M_DEVBUF);
545 		sc->csb_gh = NULL;
546 		sc->csb_hg = NULL;
547 	}
548 
549 	if (sc->queues) {
550 		for (i = 0; i < sc->num_rings; i++) {
551 			struct ptnet_queue *pq = sc->queues + i;
552 
553 			if (mtx_initialized(&pq->lock)) {
554 				mtx_destroy(&pq->lock);
555 			}
556 			if (pq->bufring != NULL) {
557 				buf_ring_free(pq->bufring, M_DEVBUF);
558 			}
559 		}
560 		free(sc->queues, M_DEVBUF);
561 		sc->queues = NULL;
562 	}
563 
564 	if (sc->iomem) {
565 		bus_release_resource(dev, SYS_RES_IOPORT,
566 				     PCIR_BAR(PTNETMAP_IO_PCI_BAR), sc->iomem);
567 		sc->iomem = NULL;
568 	}
569 
570 	mtx_destroy(&sc->lock);
571 
572 	device_printf(dev, "%s() completed\n", __func__);
573 
574 	return (0);
575 }
576 
577 static int
ptnet_suspend(device_t dev)578 ptnet_suspend(device_t dev)
579 {
580 	struct ptnet_softc *sc = device_get_softc(dev);
581 
582 	(void)sc;
583 
584 	return (0);
585 }
586 
587 static int
ptnet_resume(device_t dev)588 ptnet_resume(device_t dev)
589 {
590 	struct ptnet_softc *sc = device_get_softc(dev);
591 
592 	(void)sc;
593 
594 	return (0);
595 }
596 
597 static int
ptnet_shutdown(device_t dev)598 ptnet_shutdown(device_t dev)
599 {
600 	struct ptnet_softc *sc = device_get_softc(dev);
601 
602 	ptnet_device_shutdown(sc);
603 
604 	return (0);
605 }
606 
607 static int
ptnet_irqs_init(struct ptnet_softc * sc)608 ptnet_irqs_init(struct ptnet_softc *sc)
609 {
610 	int rid = PCIR_BAR(PTNETMAP_MSIX_PCI_BAR);
611 	int nvecs = sc->num_rings;
612 	device_t dev = sc->dev;
613 	int err = ENOSPC;
614 	int cpu_cur;
615 	int i;
616 
617 	if (pci_find_cap(dev, PCIY_MSIX, NULL) != 0)  {
618 		device_printf(dev, "Could not find MSI-X capability\n");
619 		return (ENXIO);
620 	}
621 
622 	sc->msix_mem = bus_alloc_resource_any(dev, SYS_RES_MEMORY,
623 					      &rid, RF_ACTIVE);
624 	if (sc->msix_mem == NULL) {
625 		device_printf(dev, "Failed to allocate MSIX PCI BAR\n");
626 		return (ENXIO);
627 	}
628 
629 	if (pci_msix_count(dev) < nvecs) {
630 		device_printf(dev, "Not enough MSI-X vectors\n");
631 		goto err_path;
632 	}
633 
634 	err = pci_alloc_msix(dev, &nvecs);
635 	if (err) {
636 		device_printf(dev, "Failed to allocate MSI-X vectors\n");
637 		goto err_path;
638 	}
639 
640 	for (i = 0; i < nvecs; i++) {
641 		struct ptnet_queue *pq = sc->queues + i;
642 
643 		rid = i + 1;
644 		pq->irq = bus_alloc_resource_any(dev, SYS_RES_IRQ, &rid,
645 						 RF_ACTIVE);
646 		if (pq->irq == NULL) {
647 			device_printf(dev, "Failed to allocate interrupt "
648 					   "for queue #%d\n", i);
649 			err = ENOSPC;
650 			goto err_path;
651 		}
652 	}
653 
654 	cpu_cur = CPU_FIRST();
655 	for (i = 0; i < nvecs; i++) {
656 		struct ptnet_queue *pq = sc->queues + i;
657 		void (*handler)(void *) = ptnet_tx_intr;
658 
659 		if (i >= sc->num_tx_rings) {
660 			handler = ptnet_rx_intr;
661 		}
662 		err = bus_setup_intr(dev, pq->irq, INTR_TYPE_NET | INTR_MPSAFE,
663 				     NULL /* intr_filter */, handler,
664 				     pq, &pq->cookie);
665 		if (err) {
666 			device_printf(dev, "Failed to register intr handler "
667 					   "for queue #%d\n", i);
668 			goto err_path;
669 		}
670 
671 		bus_describe_intr(dev, pq->irq, pq->cookie, "q%d", i);
672 #if 0
673 		bus_bind_intr(sc->dev, pq->irq, cpu_cur);
674 #endif
675 		cpu_cur = CPU_NEXT(cpu_cur);
676 	}
677 
678 	device_printf(dev, "Allocated %d MSI-X vectors\n", nvecs);
679 
680 	cpu_cur = CPU_FIRST();
681 	for (i = 0; i < nvecs; i++) {
682 		struct ptnet_queue *pq = sc->queues + i;
683 
684 		if (i < sc->num_tx_rings)
685 			TASK_INIT(&pq->task, 0, ptnet_tx_task, pq);
686 		else
687 			NET_TASK_INIT(&pq->task, 0, ptnet_rx_task, pq);
688 
689 		pq->taskq = taskqueue_create_fast("ptnet_queue", M_NOWAIT,
690 					taskqueue_thread_enqueue, &pq->taskq);
691 		taskqueue_start_threads(&pq->taskq, 1, PI_NET, "%s-pq-%d",
692 					device_get_nameunit(sc->dev), cpu_cur);
693 		cpu_cur = CPU_NEXT(cpu_cur);
694 	}
695 
696 	return 0;
697 err_path:
698 	ptnet_irqs_fini(sc);
699 	return err;
700 }
701 
702 static void
ptnet_irqs_fini(struct ptnet_softc * sc)703 ptnet_irqs_fini(struct ptnet_softc *sc)
704 {
705 	device_t dev = sc->dev;
706 	int i;
707 
708 	for (i = 0; i < sc->num_rings; i++) {
709 		struct ptnet_queue *pq = sc->queues + i;
710 
711 		if (pq->taskq) {
712 			taskqueue_free(pq->taskq);
713 			pq->taskq = NULL;
714 		}
715 
716 		if (pq->cookie) {
717 			bus_teardown_intr(dev, pq->irq, pq->cookie);
718 			pq->cookie = NULL;
719 		}
720 
721 		if (pq->irq) {
722 			bus_release_resource(dev, SYS_RES_IRQ, i + 1, pq->irq);
723 			pq->irq = NULL;
724 		}
725 	}
726 
727 	if (sc->msix_mem) {
728 		pci_release_msi(dev);
729 
730 		bus_release_resource(dev, SYS_RES_MEMORY,
731 				     PCIR_BAR(PTNETMAP_MSIX_PCI_BAR),
732 				     sc->msix_mem);
733 		sc->msix_mem = NULL;
734 	}
735 }
736 
737 static void
ptnet_init(void * opaque)738 ptnet_init(void *opaque)
739 {
740 	struct ptnet_softc *sc = opaque;
741 
742 	PTNET_CORE_LOCK(sc);
743 	ptnet_init_locked(sc);
744 	PTNET_CORE_UNLOCK(sc);
745 }
746 
747 static int
ptnet_ioctl(if_t ifp,u_long cmd,caddr_t data)748 ptnet_ioctl(if_t ifp, u_long cmd, caddr_t data)
749 {
750 	struct ptnet_softc *sc = if_getsoftc(ifp);
751 	device_t dev = sc->dev;
752 	struct ifreq *ifr = (struct ifreq *)data;
753 	int mask __unused, err = 0;
754 
755 	switch (cmd) {
756 	case SIOCSIFFLAGS:
757 		device_printf(dev, "SIOCSIFFLAGS %x\n", ifp->if_flags);
758 		PTNET_CORE_LOCK(sc);
759 		if (ifp->if_flags & IFF_UP) {
760 			/* Network stack wants the iff to be up. */
761 			err = ptnet_init_locked(sc);
762 		} else {
763 			/* Network stack wants the iff to be down. */
764 			err = ptnet_stop(sc);
765 		}
766 		/* We don't need to do nothing to support IFF_PROMISC,
767 		 * since that is managed by the backend port. */
768 		PTNET_CORE_UNLOCK(sc);
769 		break;
770 
771 	case SIOCSIFCAP:
772 		device_printf(dev, "SIOCSIFCAP %x %x\n",
773 			      ifr->ifr_reqcap, ifp->if_capenable);
774 		mask = ifr->ifr_reqcap ^ ifp->if_capenable;
775 #ifdef DEVICE_POLLING
776 		if (mask & IFCAP_POLLING) {
777 			struct ptnet_queue *pq;
778 			int i;
779 
780 			if (ifr->ifr_reqcap & IFCAP_POLLING) {
781 				err = ether_poll_register(ptnet_poll, ifp);
782 				if (err) {
783 					break;
784 				}
785 				/* Stop queues and sync with taskqueues. */
786 				ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
787 				for (i = 0; i < sc->num_rings; i++) {
788 					pq = sc-> queues + i;
789 					/* Make sure the worker sees the
790 					 * IFF_DRV_RUNNING down. */
791 					PTNET_Q_LOCK(pq);
792 					pq->atok->appl_need_kick = 0;
793 					PTNET_Q_UNLOCK(pq);
794 					/* Wait for rescheduling to finish. */
795 					if (pq->taskq) {
796 						taskqueue_drain(pq->taskq,
797 								&pq->task);
798 					}
799 				}
800 				ifp->if_drv_flags |= IFF_DRV_RUNNING;
801 			} else {
802 				err = ether_poll_deregister(ifp);
803 				for (i = 0; i < sc->num_rings; i++) {
804 					pq = sc-> queues + i;
805 					PTNET_Q_LOCK(pq);
806 					pq->atok->appl_need_kick = 1;
807 					PTNET_Q_UNLOCK(pq);
808 				}
809 			}
810 		}
811 #endif  /* DEVICE_POLLING */
812 		ifp->if_capenable = ifr->ifr_reqcap;
813 		break;
814 
815 	case SIOCSIFMTU:
816 		/* We support any reasonable MTU. */
817 		if (ifr->ifr_mtu < ETHERMIN ||
818 				ifr->ifr_mtu > PTNET_MAX_PKT_SIZE) {
819 			err = EINVAL;
820 		} else {
821 			PTNET_CORE_LOCK(sc);
822 			ifp->if_mtu = ifr->ifr_mtu;
823 			PTNET_CORE_UNLOCK(sc);
824 		}
825 		break;
826 
827 	case SIOCSIFMEDIA:
828 	case SIOCGIFMEDIA:
829 		err = ifmedia_ioctl(ifp, ifr, &sc->media, cmd);
830 		break;
831 
832 	default:
833 		err = ether_ioctl(ifp, cmd, data);
834 		break;
835 	}
836 
837 	return err;
838 }
839 
840 static int
ptnet_init_locked(struct ptnet_softc * sc)841 ptnet_init_locked(struct ptnet_softc *sc)
842 {
843 	if_t ifp = sc->ifp;
844 	struct netmap_adapter *na_dr = &sc->ptna->dr.up;
845 	struct netmap_adapter *na_nm = &sc->ptna->hwup.up;
846 	unsigned int nm_buf_size;
847 	int ret;
848 
849 	if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
850 		return 0; /* nothing to do */
851 	}
852 
853 	device_printf(sc->dev, "%s\n", __func__);
854 
855 	/* Translate offload capabilities according to if_capenable. */
856 	ifp->if_hwassist = 0;
857 	if (ifp->if_capenable & IFCAP_TXCSUM)
858 		ifp->if_hwassist |= PTNET_CSUM_OFFLOAD;
859 	if (ifp->if_capenable & IFCAP_TXCSUM_IPV6)
860 		ifp->if_hwassist |= PTNET_CSUM_OFFLOAD_IPV6;
861 	if (ifp->if_capenable & IFCAP_TSO4)
862 		ifp->if_hwassist |= CSUM_IP_TSO;
863 	if (ifp->if_capenable & IFCAP_TSO6)
864 		ifp->if_hwassist |= CSUM_IP6_TSO;
865 
866 	/*
867 	 * Prepare the interface for netmap mode access.
868 	 */
869 	netmap_update_config(na_dr);
870 
871 	ret = netmap_mem_finalize(na_dr->nm_mem, na_dr);
872 	if (ret) {
873 		device_printf(sc->dev, "netmap_mem_finalize() failed\n");
874 		return ret;
875 	}
876 
877 	if (sc->ptna->backend_users == 0) {
878 		ret = ptnet_nm_krings_create(na_nm);
879 		if (ret) {
880 			device_printf(sc->dev, "ptnet_nm_krings_create() "
881 					       "failed\n");
882 			goto err_mem_finalize;
883 		}
884 
885 		ret = netmap_mem_rings_create(na_dr);
886 		if (ret) {
887 			device_printf(sc->dev, "netmap_mem_rings_create() "
888 					       "failed\n");
889 			goto err_rings_create;
890 		}
891 
892 		ret = netmap_mem_get_lut(na_dr->nm_mem, &na_dr->na_lut);
893 		if (ret) {
894 			device_printf(sc->dev, "netmap_mem_get_lut() "
895 					       "failed\n");
896 			goto err_get_lut;
897 		}
898 	}
899 
900 	ret = ptnet_nm_register(na_dr, 1 /* on */);
901 	if (ret) {
902 		goto err_register;
903 	}
904 
905 	nm_buf_size = NETMAP_BUF_SIZE(na_dr);
906 
907 	KASSERT(nm_buf_size > 0, ("Invalid netmap buffer size"));
908 	sc->min_tx_space = PTNET_MAX_PKT_SIZE / nm_buf_size + 2;
909 	device_printf(sc->dev, "%s: min_tx_space = %u\n", __func__,
910 		      sc->min_tx_space);
911 #ifdef PTNETMAP_STATS
912 	callout_reset(&sc->tick, hz, ptnet_tick, sc);
913 #endif
914 
915 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
916 
917 	return 0;
918 
919 err_register:
920 	memset(&na_dr->na_lut, 0, sizeof(na_dr->na_lut));
921 err_get_lut:
922 	netmap_mem_rings_delete(na_dr);
923 err_rings_create:
924 	ptnet_nm_krings_delete(na_nm);
925 err_mem_finalize:
926 	netmap_mem_deref(na_dr->nm_mem, na_dr);
927 
928 	return ret;
929 }
930 
931 /* To be called under core lock. */
932 static int
ptnet_stop(struct ptnet_softc * sc)933 ptnet_stop(struct ptnet_softc *sc)
934 {
935 	if_t ifp = sc->ifp;
936 	struct netmap_adapter *na_dr = &sc->ptna->dr.up;
937 	struct netmap_adapter *na_nm = &sc->ptna->hwup.up;
938 	int i;
939 
940 	device_printf(sc->dev, "%s\n", __func__);
941 
942 	if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) {
943 		return 0; /* nothing to do */
944 	}
945 
946 	/* Clear the driver-ready flag, and synchronize with all the queues,
947 	 * so that after this loop we are sure nobody is working anymore with
948 	 * the device. This scheme is taken from the vtnet driver. */
949 	ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
950 	callout_stop(&sc->tick);
951 	for (i = 0; i < sc->num_rings; i++) {
952 		PTNET_Q_LOCK(sc->queues + i);
953 		PTNET_Q_UNLOCK(sc->queues + i);
954 	}
955 
956 	ptnet_nm_register(na_dr, 0 /* off */);
957 
958 	if (sc->ptna->backend_users == 0) {
959 		netmap_mem_rings_delete(na_dr);
960 		ptnet_nm_krings_delete(na_nm);
961 	}
962 	netmap_mem_deref(na_dr->nm_mem, na_dr);
963 
964 	return 0;
965 }
966 
967 static void
ptnet_qflush(if_t ifp)968 ptnet_qflush(if_t ifp)
969 {
970 	struct ptnet_softc *sc = if_getsoftc(ifp);
971 	int i;
972 
973 	/* Flush all the bufrings and do the interface flush. */
974 	for (i = 0; i < sc->num_rings; i++) {
975 		struct ptnet_queue *pq = sc->queues + i;
976 		struct mbuf *m;
977 
978 		PTNET_Q_LOCK(pq);
979 		if (pq->bufring) {
980 			while ((m = buf_ring_dequeue_sc(pq->bufring))) {
981 				m_freem(m);
982 			}
983 		}
984 		PTNET_Q_UNLOCK(pq);
985 	}
986 
987 	if_qflush(ifp);
988 }
989 
990 static int
ptnet_media_change(if_t ifp)991 ptnet_media_change(if_t ifp)
992 {
993 	struct ptnet_softc *sc = if_getsoftc(ifp);
994 	struct ifmedia *ifm = &sc->media;
995 
996 	if (IFM_TYPE(ifm->ifm_media) != IFM_ETHER) {
997 		return EINVAL;
998 	}
999 
1000 	return 0;
1001 }
1002 
1003 static uint64_t
ptnet_get_counter(if_t ifp,ift_counter cnt)1004 ptnet_get_counter(if_t ifp, ift_counter cnt)
1005 {
1006 	struct ptnet_softc *sc = if_getsoftc(ifp);
1007 	struct ptnet_queue_stats stats[2];
1008 	int i;
1009 
1010 	/* Accumulate statistics over the queues. */
1011 	memset(stats, 0, sizeof(stats));
1012 	for (i = 0; i < sc->num_rings; i++) {
1013 		struct ptnet_queue *pq = sc->queues + i;
1014 		int idx = (i < sc->num_tx_rings) ? 0 : 1;
1015 
1016 		stats[idx].packets	+= pq->stats.packets;
1017 		stats[idx].bytes	+= pq->stats.bytes;
1018 		stats[idx].errors	+= pq->stats.errors;
1019 		stats[idx].iqdrops	+= pq->stats.iqdrops;
1020 		stats[idx].mcasts	+= pq->stats.mcasts;
1021 	}
1022 
1023 	switch (cnt) {
1024 	case IFCOUNTER_IPACKETS:
1025 		return (stats[1].packets);
1026 	case IFCOUNTER_IQDROPS:
1027 		return (stats[1].iqdrops);
1028 	case IFCOUNTER_IERRORS:
1029 		return (stats[1].errors);
1030 	case IFCOUNTER_OPACKETS:
1031 		return (stats[0].packets);
1032 	case IFCOUNTER_OBYTES:
1033 		return (stats[0].bytes);
1034 	case IFCOUNTER_OMCASTS:
1035 		return (stats[0].mcasts);
1036 	default:
1037 		return (if_get_counter_default(ifp, cnt));
1038 	}
1039 }
1040 
1041 
1042 #ifdef PTNETMAP_STATS
1043 /* Called under core lock. */
1044 static void
ptnet_tick(void * opaque)1045 ptnet_tick(void *opaque)
1046 {
1047 	struct ptnet_softc *sc = opaque;
1048 	int i;
1049 
1050 	for (i = 0; i < sc->num_rings; i++) {
1051 		struct ptnet_queue *pq = sc->queues + i;
1052 		struct ptnet_queue_stats cur = pq->stats;
1053 		struct timeval now;
1054 		unsigned int delta;
1055 
1056 		microtime(&now);
1057 		delta = now.tv_usec - sc->last_ts.tv_usec +
1058 			(now.tv_sec - sc->last_ts.tv_sec) * 1000000;
1059 		delta /= 1000; /* in milliseconds */
1060 
1061 		if (delta == 0)
1062 			continue;
1063 
1064 		device_printf(sc->dev, "#%d[%u ms]:pkts %lu, kicks %lu, "
1065 			      "intr %lu\n", i, delta,
1066 			      (cur.packets - pq->last_stats.packets),
1067 			      (cur.kicks - pq->last_stats.kicks),
1068 			      (cur.intrs - pq->last_stats.intrs));
1069 		pq->last_stats = cur;
1070 	}
1071 	microtime(&sc->last_ts);
1072 	callout_schedule(&sc->tick, hz);
1073 }
1074 #endif /* PTNETMAP_STATS */
1075 
1076 static void
ptnet_media_status(if_t ifp,struct ifmediareq * ifmr)1077 ptnet_media_status(if_t ifp, struct ifmediareq *ifmr)
1078 {
1079 	/* We are always active, as the backend netmap port is
1080 	 * always open in netmap mode. */
1081 	ifmr->ifm_status = IFM_AVALID | IFM_ACTIVE;
1082 	ifmr->ifm_active = IFM_ETHER | IFM_10G_T | IFM_FDX;
1083 }
1084 
1085 static uint32_t
ptnet_nm_ptctl(struct ptnet_softc * sc,uint32_t cmd)1086 ptnet_nm_ptctl(struct ptnet_softc *sc, uint32_t cmd)
1087 {
1088 	/*
1089 	 * Write a command and read back error status,
1090 	 * with zero meaning success.
1091 	 */
1092 	bus_write_4(sc->iomem, PTNET_IO_PTCTL, cmd);
1093 	return bus_read_4(sc->iomem, PTNET_IO_PTCTL);
1094 }
1095 
1096 static int
ptnet_nm_config(struct netmap_adapter * na,struct nm_config_info * info)1097 ptnet_nm_config(struct netmap_adapter *na, struct nm_config_info *info)
1098 {
1099 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1100 
1101 	info->num_tx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_RINGS);
1102 	info->num_rx_rings = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_RINGS);
1103 	info->num_tx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_TX_SLOTS);
1104 	info->num_rx_descs = bus_read_4(sc->iomem, PTNET_IO_NUM_RX_SLOTS);
1105 	info->rx_buf_maxsize = NETMAP_BUF_SIZE(na);
1106 
1107 	device_printf(sc->dev, "txr %u, rxr %u, txd %u, rxd %u, rxbufsz %u\n",
1108 			info->num_tx_rings, info->num_rx_rings,
1109 			info->num_tx_descs, info->num_rx_descs,
1110 			info->rx_buf_maxsize);
1111 
1112 	return 0;
1113 }
1114 
1115 static void
ptnet_sync_from_csb(struct ptnet_softc * sc,struct netmap_adapter * na)1116 ptnet_sync_from_csb(struct ptnet_softc *sc, struct netmap_adapter *na)
1117 {
1118 	int i;
1119 
1120 	/* Sync krings from the host, reading from
1121 	 * CSB. */
1122 	for (i = 0; i < sc->num_rings; i++) {
1123 		struct nm_csb_atok *atok = sc->queues[i].atok;
1124 		struct nm_csb_ktoa *ktoa = sc->queues[i].ktoa;
1125 		struct netmap_kring *kring;
1126 
1127 		if (i < na->num_tx_rings) {
1128 			kring = na->tx_rings[i];
1129 		} else {
1130 			kring = na->rx_rings[i - na->num_tx_rings];
1131 		}
1132 		kring->rhead = kring->ring->head = atok->head;
1133 		kring->rcur = kring->ring->cur = atok->cur;
1134 		kring->nr_hwcur = ktoa->hwcur;
1135 		kring->nr_hwtail = kring->rtail =
1136 			kring->ring->tail = ktoa->hwtail;
1137 
1138 		nm_prdis("%d,%d: csb {hc %u h %u c %u ht %u}", t, i,
1139 		   ktoa->hwcur, atok->head, atok->cur,
1140 		   ktoa->hwtail);
1141 		nm_prdis("%d,%d: kring {hc %u rh %u rc %u h %u c %u ht %u rt %u t %u}",
1142 		   t, i, kring->nr_hwcur, kring->rhead, kring->rcur,
1143 		   kring->ring->head, kring->ring->cur, kring->nr_hwtail,
1144 		   kring->rtail, kring->ring->tail);
1145 	}
1146 }
1147 
1148 static void
ptnet_update_vnet_hdr(struct ptnet_softc * sc)1149 ptnet_update_vnet_hdr(struct ptnet_softc *sc)
1150 {
1151 	unsigned int wanted_hdr_len = ptnet_vnet_hdr ? PTNET_HDR_SIZE : 0;
1152 
1153 	bus_write_4(sc->iomem, PTNET_IO_VNET_HDR_LEN, wanted_hdr_len);
1154 	sc->vnet_hdr_len = bus_read_4(sc->iomem, PTNET_IO_VNET_HDR_LEN);
1155 	sc->ptna->hwup.up.virt_hdr_len = sc->vnet_hdr_len;
1156 }
1157 
1158 static int
ptnet_nm_register(struct netmap_adapter * na,int onoff)1159 ptnet_nm_register(struct netmap_adapter *na, int onoff)
1160 {
1161 	/* device-specific */
1162 	if_t ifp = na->ifp;
1163 	struct ptnet_softc *sc = if_getsoftc(ifp);
1164 	int native = (na == &sc->ptna->hwup.up);
1165 	struct ptnet_queue *pq;
1166 	int ret = 0;
1167 	int i;
1168 
1169 	if (!onoff) {
1170 		sc->ptna->backend_users--;
1171 	}
1172 
1173 	/* If this is the last netmap client, guest interrupt enable flags may
1174 	 * be in arbitrary state. Since these flags are going to be used also
1175 	 * by the netdevice driver, we have to make sure to start with
1176 	 * notifications enabled. Also, schedule NAPI to flush pending packets
1177 	 * in the RX rings, since we will not receive further interrupts
1178 	 * until these will be processed. */
1179 	if (native && !onoff && na->active_fds == 0) {
1180 		nm_prinf("Exit netmap mode, re-enable interrupts");
1181 		for (i = 0; i < sc->num_rings; i++) {
1182 			pq = sc->queues + i;
1183 			pq->atok->appl_need_kick = 1;
1184 		}
1185 	}
1186 
1187 	if (onoff) {
1188 		if (sc->ptna->backend_users == 0) {
1189 			/* Initialize notification enable fields in the CSB. */
1190 			for (i = 0; i < sc->num_rings; i++) {
1191 				pq = sc->queues + i;
1192 				pq->ktoa->kern_need_kick = 1;
1193 				pq->atok->appl_need_kick =
1194 					(!(ifp->if_capenable & IFCAP_POLLING)
1195 						&& i >= sc->num_tx_rings);
1196 			}
1197 
1198 			/* Set the virtio-net header length. */
1199 			ptnet_update_vnet_hdr(sc);
1200 
1201 			/* Make sure the host adapter passed through is ready
1202 			 * for txsync/rxsync. */
1203 			ret = ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_CREATE);
1204 			if (ret) {
1205 				return ret;
1206 			}
1207 
1208 			/* Align the guest krings and rings to the state stored
1209 			 * in the CSB. */
1210 			ptnet_sync_from_csb(sc, na);
1211 		}
1212 
1213 		/* If not native, don't call nm_set_native_flags, since we don't want
1214 		 * to replace if_transmit method, nor set NAF_NETMAP_ON */
1215 		if (native) {
1216 			netmap_krings_mode_commit(na, onoff);
1217 			nm_set_native_flags(na);
1218 		}
1219 
1220 	} else {
1221 		if (native) {
1222 			nm_clear_native_flags(na);
1223 			netmap_krings_mode_commit(na, onoff);
1224 		}
1225 
1226 		if (sc->ptna->backend_users == 0) {
1227 			ret = ptnet_nm_ptctl(sc, PTNETMAP_PTCTL_DELETE);
1228 		}
1229 	}
1230 
1231 	if (onoff) {
1232 		sc->ptna->backend_users++;
1233 	}
1234 
1235 	return ret;
1236 }
1237 
1238 static int
ptnet_nm_txsync(struct netmap_kring * kring,int flags)1239 ptnet_nm_txsync(struct netmap_kring *kring, int flags)
1240 {
1241 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1242 	struct ptnet_queue *pq = sc->queues + kring->ring_id;
1243 	bool notify;
1244 
1245 	notify = netmap_pt_guest_txsync(pq->atok, pq->ktoa, kring, flags);
1246 	if (notify) {
1247 		ptnet_kick(pq);
1248 	}
1249 
1250 	return 0;
1251 }
1252 
1253 static int
ptnet_nm_rxsync(struct netmap_kring * kring,int flags)1254 ptnet_nm_rxsync(struct netmap_kring *kring, int flags)
1255 {
1256 	struct ptnet_softc *sc = if_getsoftc(kring->na->ifp);
1257 	struct ptnet_queue *pq = sc->rxqueues + kring->ring_id;
1258 	bool notify;
1259 
1260 	notify = netmap_pt_guest_rxsync(pq->atok, pq->ktoa, kring, flags);
1261 	if (notify) {
1262 		ptnet_kick(pq);
1263 	}
1264 
1265 	return 0;
1266 }
1267 
1268 static void
ptnet_nm_intr(struct netmap_adapter * na,int onoff)1269 ptnet_nm_intr(struct netmap_adapter *na, int onoff)
1270 {
1271 	struct ptnet_softc *sc = if_getsoftc(na->ifp);
1272 	int i;
1273 
1274 	for (i = 0; i < sc->num_rings; i++) {
1275 		struct ptnet_queue *pq = sc->queues + i;
1276 		pq->atok->appl_need_kick = onoff;
1277 	}
1278 }
1279 
1280 static void
ptnet_tx_intr(void * opaque)1281 ptnet_tx_intr(void *opaque)
1282 {
1283 	struct ptnet_queue *pq = opaque;
1284 	struct ptnet_softc *sc = pq->sc;
1285 
1286 	DBG(device_printf(sc->dev, "Tx interrupt #%d\n", pq->kring_id));
1287 #ifdef PTNETMAP_STATS
1288 	pq->stats.intrs ++;
1289 #endif /* PTNETMAP_STATS */
1290 
1291 	if (netmap_tx_irq(sc->ifp, pq->kring_id) != NM_IRQ_PASS) {
1292 		return;
1293 	}
1294 
1295 	/* Schedule the tasqueue to flush process transmissions requests.
1296 	 * However, vtnet, if_em and if_igb just call ptnet_transmit() here,
1297 	 * at least when using MSI-X interrupts. The if_em driver, instead
1298 	 * schedule taskqueue when using legacy interrupts. */
1299 	taskqueue_enqueue(pq->taskq, &pq->task);
1300 }
1301 
1302 static void
ptnet_rx_intr(void * opaque)1303 ptnet_rx_intr(void *opaque)
1304 {
1305 	struct ptnet_queue *pq = opaque;
1306 	struct ptnet_softc *sc = pq->sc;
1307 	unsigned int unused;
1308 
1309 	DBG(device_printf(sc->dev, "Rx interrupt #%d\n", pq->kring_id));
1310 #ifdef PTNETMAP_STATS
1311 	pq->stats.intrs ++;
1312 #endif /* PTNETMAP_STATS */
1313 
1314 	if (netmap_rx_irq(sc->ifp, pq->kring_id, &unused) != NM_IRQ_PASS) {
1315 		return;
1316 	}
1317 
1318 	/* Like vtnet, if_igb and if_em drivers when using MSI-X interrupts,
1319 	 * receive-side processing is executed directly in the interrupt
1320 	 * service routine. Alternatively, we may schedule the taskqueue. */
1321 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
1322 }
1323 
1324 static void
ptnet_vlan_tag_remove(struct mbuf * m)1325 ptnet_vlan_tag_remove(struct mbuf *m)
1326 {
1327 	struct ether_vlan_header *evh;
1328 
1329 	evh = mtod(m, struct ether_vlan_header *);
1330 	m->m_pkthdr.ether_vtag = ntohs(evh->evl_tag);
1331 	m->m_flags |= M_VLANTAG;
1332 
1333 	/* Strip the 802.1Q header. */
1334 	bcopy((char *) evh, (char *) evh + ETHER_VLAN_ENCAP_LEN,
1335 	    ETHER_HDR_LEN - ETHER_TYPE_LEN);
1336 	m_adj(m, ETHER_VLAN_ENCAP_LEN);
1337 }
1338 
1339 static void
ptnet_ring_update(struct ptnet_queue * pq,struct netmap_kring * kring,unsigned int head,unsigned int sync_flags)1340 ptnet_ring_update(struct ptnet_queue *pq, struct netmap_kring *kring,
1341 		  unsigned int head, unsigned int sync_flags)
1342 {
1343 	struct netmap_ring *ring = kring->ring;
1344 	struct nm_csb_atok *atok = pq->atok;
1345 	struct nm_csb_ktoa *ktoa = pq->ktoa;
1346 
1347 	/* Some packets have been pushed to the netmap ring. We have
1348 	 * to tell the host to process the new packets, updating cur
1349 	 * and head in the CSB. */
1350 	ring->head = ring->cur = head;
1351 
1352 	/* Mimic nm_txsync_prologue/nm_rxsync_prologue. */
1353 	kring->rcur = kring->rhead = head;
1354 
1355 	nm_sync_kloop_appl_write(atok, kring->rcur, kring->rhead);
1356 
1357 	/* Kick the host if needed. */
1358 	if (NM_ACCESS_ONCE(ktoa->kern_need_kick)) {
1359 		atok->sync_flags = sync_flags;
1360 		ptnet_kick(pq);
1361 	}
1362 }
1363 
1364 #define PTNET_TX_NOSPACE(_h, _k, _min)	\
1365 	((((_h) < (_k)->rtail) ? 0 : (_k)->nkr_num_slots) + \
1366 		(_k)->rtail - (_h)) < (_min)
1367 
1368 /* This function may be called by the network stack, or by
1369  * by the taskqueue thread. */
1370 static int
ptnet_drain_transmit_queue(struct ptnet_queue * pq,unsigned int budget,bool may_resched)1371 ptnet_drain_transmit_queue(struct ptnet_queue *pq, unsigned int budget,
1372 			   bool may_resched)
1373 {
1374 	struct ptnet_softc *sc = pq->sc;
1375 	bool have_vnet_hdr = sc->vnet_hdr_len;
1376 	struct netmap_adapter *na = &sc->ptna->dr.up;
1377 	if_t ifp = sc->ifp;
1378 	unsigned int batch_count = 0;
1379 	struct nm_csb_atok *atok;
1380 	struct nm_csb_ktoa *ktoa;
1381 	struct netmap_kring *kring;
1382 	struct netmap_ring *ring;
1383 	struct netmap_slot *slot;
1384 	unsigned int count = 0;
1385 	unsigned int minspace;
1386 	unsigned int head;
1387 	unsigned int lim;
1388 	struct mbuf *mhead;
1389 	struct mbuf *mf;
1390 	int nmbuf_bytes;
1391 	uint8_t *nmbuf;
1392 
1393 	if (!PTNET_Q_TRYLOCK(pq)) {
1394 		/* We failed to acquire the lock, schedule the taskqueue. */
1395 		nm_prlim(1, "Deferring TX work");
1396 		if (may_resched) {
1397 			taskqueue_enqueue(pq->taskq, &pq->task);
1398 		}
1399 
1400 		return 0;
1401 	}
1402 
1403 	if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1404 		PTNET_Q_UNLOCK(pq);
1405 		nm_prlim(1, "Interface is down");
1406 		return ENETDOWN;
1407 	}
1408 
1409 	atok = pq->atok;
1410 	ktoa = pq->ktoa;
1411 	kring = na->tx_rings[pq->kring_id];
1412 	ring = kring->ring;
1413 	lim = kring->nkr_num_slots - 1;
1414 	head = ring->head;
1415 	minspace = sc->min_tx_space;
1416 
1417 	while (count < budget) {
1418 		if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1419 			/* We ran out of slot, let's see if the host has
1420 			 * freed up some, by reading hwcur and hwtail from
1421 			 * the CSB. */
1422 			ptnet_sync_tail(ktoa, kring);
1423 
1424 			if (PTNET_TX_NOSPACE(head, kring, minspace)) {
1425 				/* Still no slots available. Reactivate the
1426 				 * interrupts so that we can be notified
1427 				 * when some free slots are made available by
1428 				 * the host. */
1429 				atok->appl_need_kick = 1;
1430 
1431 				/* Double check. We need a full barrier to
1432 				 * prevent the store to atok->appl_need_kick
1433 				 * to be reordered with the load from
1434 				 * ktoa->hwcur and ktoa->hwtail (store-load
1435 				 * barrier). */
1436 				nm_stld_barrier();
1437 				ptnet_sync_tail(ktoa, kring);
1438 				if (likely(PTNET_TX_NOSPACE(head, kring,
1439 							    minspace))) {
1440 					break;
1441 				}
1442 
1443 				nm_prlim(1, "Found more slots by doublecheck");
1444 				/* More slots were freed before reactivating
1445 				 * the interrupts. */
1446 				atok->appl_need_kick = 0;
1447 			}
1448 		}
1449 
1450 		mhead = drbr_peek(ifp, pq->bufring);
1451 		if (!mhead) {
1452 			break;
1453 		}
1454 
1455 		/* Initialize transmission state variables. */
1456 		slot = ring->slot + head;
1457 		nmbuf = NMB(na, slot);
1458 		nmbuf_bytes = 0;
1459 
1460 		/* If needed, prepare the virtio-net header at the beginning
1461 		 * of the first slot. */
1462 		if (have_vnet_hdr) {
1463 			struct virtio_net_hdr *vh =
1464 					(struct virtio_net_hdr *)nmbuf;
1465 
1466 			/* For performance, we could replace this memset() with
1467 			 * two 8-bytes-wide writes. */
1468 			memset(nmbuf, 0, PTNET_HDR_SIZE);
1469 			if (mhead->m_pkthdr.csum_flags & PTNET_ALL_OFFLOAD) {
1470 				mhead = virtio_net_tx_offload(ifp, mhead, false,
1471 							 vh);
1472 				if (unlikely(!mhead)) {
1473 					/* Packet dropped because errors
1474 					 * occurred while preparing the vnet
1475 					 * header. Let's go ahead with the next
1476 					 * packet. */
1477 					pq->stats.errors ++;
1478 					drbr_advance(ifp, pq->bufring);
1479 					continue;
1480 				}
1481 			}
1482 			nm_prdis(1, "%s: [csum_flags %lX] vnet hdr: flags %x "
1483 			      "csum_start %u csum_ofs %u hdr_len = %u "
1484 			      "gso_size %u gso_type %x", __func__,
1485 			      mhead->m_pkthdr.csum_flags, vh->flags,
1486 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
1487 			      vh->gso_size, vh->gso_type);
1488 
1489 			nmbuf += PTNET_HDR_SIZE;
1490 			nmbuf_bytes += PTNET_HDR_SIZE;
1491 		}
1492 
1493 		for (mf = mhead; mf; mf = mf->m_next) {
1494 			uint8_t *mdata = mf->m_data;
1495 			int mlen = mf->m_len;
1496 
1497 			for (;;) {
1498 				int copy = NETMAP_BUF_SIZE(na) - nmbuf_bytes;
1499 
1500 				if (mlen < copy) {
1501 					copy = mlen;
1502 				}
1503 				memcpy(nmbuf, mdata, copy);
1504 
1505 				mdata += copy;
1506 				mlen -= copy;
1507 				nmbuf += copy;
1508 				nmbuf_bytes += copy;
1509 
1510 				if (!mlen) {
1511 					break;
1512 				}
1513 
1514 				slot->len = nmbuf_bytes;
1515 				slot->flags = NS_MOREFRAG;
1516 
1517 				head = nm_next(head, lim);
1518 				KASSERT(head != ring->tail,
1519 					("Unexpectedly run out of TX space"));
1520 				slot = ring->slot + head;
1521 				nmbuf = NMB(na, slot);
1522 				nmbuf_bytes = 0;
1523 			}
1524 		}
1525 
1526 		/* Complete last slot and update head. */
1527 		slot->len = nmbuf_bytes;
1528 		slot->flags = 0;
1529 		head = nm_next(head, lim);
1530 
1531 		/* Consume the packet just processed. */
1532 		drbr_advance(ifp, pq->bufring);
1533 
1534 		/* Copy the packet to listeners. */
1535 		ETHER_BPF_MTAP(ifp, mhead);
1536 
1537 		pq->stats.packets ++;
1538 		pq->stats.bytes += mhead->m_pkthdr.len;
1539 		if (mhead->m_flags & M_MCAST) {
1540 			pq->stats.mcasts ++;
1541 		}
1542 
1543 		m_freem(mhead);
1544 
1545 		count ++;
1546 		if (++batch_count == PTNET_TX_BATCH) {
1547 			ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1548 			batch_count = 0;
1549 		}
1550 	}
1551 
1552 	if (batch_count) {
1553 		ptnet_ring_update(pq, kring, head, NAF_FORCE_RECLAIM);
1554 	}
1555 
1556 	if (count >= budget && may_resched) {
1557 		DBG(nm_prlim(1, "out of budget: resched, %d mbufs pending\n",
1558 					drbr_inuse(ifp, pq->bufring)));
1559 		taskqueue_enqueue(pq->taskq, &pq->task);
1560 	}
1561 
1562 	PTNET_Q_UNLOCK(pq);
1563 
1564 	return count;
1565 }
1566 
1567 static int
ptnet_transmit(if_t ifp,struct mbuf * m)1568 ptnet_transmit(if_t ifp, struct mbuf *m)
1569 {
1570 	struct ptnet_softc *sc = if_getsoftc(ifp);
1571 	struct ptnet_queue *pq;
1572 	unsigned int queue_idx;
1573 	int err;
1574 
1575 	DBG(device_printf(sc->dev, "transmit %p\n", m));
1576 
1577 	/* Insert 802.1Q header if needed. */
1578 	if (m->m_flags & M_VLANTAG) {
1579 		m = ether_vlanencap(m, m->m_pkthdr.ether_vtag);
1580 		if (m == NULL) {
1581 			return ENOBUFS;
1582 		}
1583 		m->m_flags &= ~M_VLANTAG;
1584 	}
1585 
1586 	/* Get the flow-id if available. */
1587 	queue_idx = (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) ?
1588 		    m->m_pkthdr.flowid : curcpu;
1589 
1590 	if (unlikely(queue_idx >= sc->num_tx_rings)) {
1591 		queue_idx %= sc->num_tx_rings;
1592 	}
1593 
1594 	pq = sc->queues + queue_idx;
1595 
1596 	err = drbr_enqueue(ifp, pq->bufring, m);
1597 	if (err) {
1598 		/* ENOBUFS when the bufring is full */
1599 		nm_prlim(1, "%s: drbr_enqueue() failed %d\n",
1600 			__func__, err);
1601 		pq->stats.errors ++;
1602 		return err;
1603 	}
1604 
1605 	if (ifp->if_capenable & IFCAP_POLLING) {
1606 		/* If polling is on, the transmit queues will be
1607 		 * drained by the poller. */
1608 		return 0;
1609 	}
1610 
1611 	err = ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
1612 
1613 	return (err < 0) ? err : 0;
1614 }
1615 
1616 static unsigned int
ptnet_rx_discard(struct netmap_kring * kring,unsigned int head)1617 ptnet_rx_discard(struct netmap_kring *kring, unsigned int head)
1618 {
1619 	struct netmap_ring *ring = kring->ring;
1620 	struct netmap_slot *slot = ring->slot + head;
1621 
1622 	for (;;) {
1623 		head = nm_next(head, kring->nkr_num_slots - 1);
1624 		if (!(slot->flags & NS_MOREFRAG) || head == ring->tail) {
1625 			break;
1626 		}
1627 		slot = ring->slot + head;
1628 	}
1629 
1630 	return head;
1631 }
1632 
1633 static inline struct mbuf *
ptnet_rx_slot(struct mbuf * mtail,uint8_t * nmbuf,unsigned int nmbuf_len)1634 ptnet_rx_slot(struct mbuf *mtail, uint8_t *nmbuf, unsigned int nmbuf_len)
1635 {
1636 	uint8_t *mdata = mtod(mtail, uint8_t *) + mtail->m_len;
1637 
1638 	do {
1639 		unsigned int copy;
1640 
1641 		if (mtail->m_len == MCLBYTES) {
1642 			struct mbuf *mf;
1643 
1644 			mf = m_getcl(M_NOWAIT, MT_DATA, 0);
1645 			if (unlikely(!mf)) {
1646 				return NULL;
1647 			}
1648 
1649 			mtail->m_next = mf;
1650 			mtail = mf;
1651 			mdata = mtod(mtail, uint8_t *);
1652 			mtail->m_len = 0;
1653 		}
1654 
1655 		copy = MCLBYTES - mtail->m_len;
1656 		if (nmbuf_len < copy) {
1657 			copy = nmbuf_len;
1658 		}
1659 
1660 		memcpy(mdata, nmbuf, copy);
1661 
1662 		nmbuf += copy;
1663 		nmbuf_len -= copy;
1664 		mdata += copy;
1665 		mtail->m_len += copy;
1666 	} while (nmbuf_len);
1667 
1668 	return mtail;
1669 }
1670 
1671 static int
ptnet_rx_eof(struct ptnet_queue * pq,unsigned int budget,bool may_resched)1672 ptnet_rx_eof(struct ptnet_queue *pq, unsigned int budget, bool may_resched)
1673 {
1674 	struct ptnet_softc *sc = pq->sc;
1675 	bool have_vnet_hdr = sc->vnet_hdr_len;
1676 	struct nm_csb_atok *atok = pq->atok;
1677 	struct nm_csb_ktoa *ktoa = pq->ktoa;
1678 	struct netmap_adapter *na = &sc->ptna->dr.up;
1679 	struct netmap_kring *kring = na->rx_rings[pq->kring_id];
1680 	struct netmap_ring *ring = kring->ring;
1681 	unsigned int const lim = kring->nkr_num_slots - 1;
1682 	unsigned int batch_count = 0;
1683 	if_t ifp = sc->ifp;
1684 	unsigned int count = 0;
1685 	uint32_t head;
1686 
1687 	PTNET_Q_LOCK(pq);
1688 
1689 	if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1690 		goto unlock;
1691 	}
1692 
1693 	kring->nr_kflags &= ~NKR_PENDINTR;
1694 
1695 	head = ring->head;
1696 	while (count < budget) {
1697 		uint32_t prev_head = head;
1698 		struct mbuf *mhead, *mtail;
1699 		struct virtio_net_hdr *vh;
1700 		struct netmap_slot *slot;
1701 		unsigned int nmbuf_len;
1702 		uint8_t *nmbuf;
1703 		int deliver = 1; /* the mbuf to the network stack. */
1704 host_sync:
1705 		if (head == ring->tail) {
1706 			/* We ran out of slot, let's see if the host has
1707 			 * added some, by reading hwcur and hwtail from
1708 			 * the CSB. */
1709 			ptnet_sync_tail(ktoa, kring);
1710 
1711 			if (head == ring->tail) {
1712 				/* Still no slots available. Reactivate
1713 				 * interrupts as they were disabled by the
1714 				 * host thread right before issuing the
1715 				 * last interrupt. */
1716 				atok->appl_need_kick = 1;
1717 
1718 				/* Double check for more completed RX slots.
1719 				 * We need a full barrier to prevent the store
1720 				 * to atok->appl_need_kick to be reordered with
1721 				 * the load from ktoa->hwcur and ktoa->hwtail
1722 				 * (store-load barrier). */
1723 				nm_stld_barrier();
1724 				ptnet_sync_tail(ktoa, kring);
1725 				if (likely(head == ring->tail)) {
1726 					break;
1727 				}
1728 				atok->appl_need_kick = 0;
1729 			}
1730 		}
1731 
1732 		/* Initialize ring state variables, possibly grabbing the
1733 		 * virtio-net header. */
1734 		slot = ring->slot + head;
1735 		nmbuf = NMB(na, slot);
1736 		nmbuf_len = slot->len;
1737 
1738 		vh = (struct virtio_net_hdr *)nmbuf;
1739 		if (have_vnet_hdr) {
1740 			if (unlikely(nmbuf_len < PTNET_HDR_SIZE)) {
1741 				/* There is no good reason why host should
1742 				 * put the header in multiple netmap slots.
1743 				 * If this is the case, discard. */
1744 				nm_prlim(1, "Fragmented vnet-hdr: dropping");
1745 				head = ptnet_rx_discard(kring, head);
1746 				pq->stats.iqdrops ++;
1747 				deliver = 0;
1748 				goto skip;
1749 			}
1750 			nm_prdis(1, "%s: vnet hdr: flags %x csum_start %u "
1751 			      "csum_ofs %u hdr_len = %u gso_size %u "
1752 			      "gso_type %x", __func__, vh->flags,
1753 			      vh->csum_start, vh->csum_offset, vh->hdr_len,
1754 			      vh->gso_size, vh->gso_type);
1755 			nmbuf += PTNET_HDR_SIZE;
1756 			nmbuf_len -= PTNET_HDR_SIZE;
1757 		}
1758 
1759 		/* Allocate the head of a new mbuf chain.
1760 		 * We use m_getcl() to allocate an mbuf with standard cluster
1761 		 * size (MCLBYTES). In the future we could use m_getjcl()
1762 		 * to choose different sizes. */
1763 		mhead = mtail = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1764 		if (unlikely(mhead == NULL)) {
1765 			device_printf(sc->dev, "%s: failed to allocate mbuf "
1766 				      "head\n", __func__);
1767 			pq->stats.errors ++;
1768 			break;
1769 		}
1770 
1771 		/* Initialize the mbuf state variables. */
1772 		mhead->m_pkthdr.len = nmbuf_len;
1773 		mtail->m_len = 0;
1774 
1775 		/* Scan all the netmap slots containing the current packet. */
1776 		for (;;) {
1777 			DBG(device_printf(sc->dev, "%s: h %u t %u rcv frag "
1778 					  "len %u, flags %u\n", __func__,
1779 					  head, ring->tail, slot->len,
1780 					  slot->flags));
1781 
1782 			mtail = ptnet_rx_slot(mtail, nmbuf, nmbuf_len);
1783 			if (unlikely(!mtail)) {
1784 				/* Ouch. We ran out of memory while processing
1785 				 * a packet. We have to restore the previous
1786 				 * head position, free the mbuf chain, and
1787 				 * schedule the taskqueue to give the packet
1788 				 * another chance. */
1789 				device_printf(sc->dev, "%s: failed to allocate"
1790 					" mbuf frag, reset head %u --> %u\n",
1791 					__func__, head, prev_head);
1792 				head = prev_head;
1793 				m_freem(mhead);
1794 				pq->stats.errors ++;
1795 				if (may_resched) {
1796 					taskqueue_enqueue(pq->taskq,
1797 							  &pq->task);
1798 				}
1799 				goto escape;
1800 			}
1801 
1802 			/* We have to increment head irrespective of the
1803 			 * NS_MOREFRAG being set or not. */
1804 			head = nm_next(head, lim);
1805 
1806 			if (!(slot->flags & NS_MOREFRAG)) {
1807 				break;
1808 			}
1809 
1810 			if (unlikely(head == ring->tail)) {
1811 				/* The very last slot prepared by the host has
1812 				 * the NS_MOREFRAG set. Drop it and continue
1813 				 * the outer cycle (to do the double-check). */
1814 				nm_prlim(1, "Incomplete packet: dropping");
1815 				m_freem(mhead);
1816 				pq->stats.iqdrops ++;
1817 				goto host_sync;
1818 			}
1819 
1820 			slot = ring->slot + head;
1821 			nmbuf = NMB(na, slot);
1822 			nmbuf_len = slot->len;
1823 			mhead->m_pkthdr.len += nmbuf_len;
1824 		}
1825 
1826 		mhead->m_pkthdr.rcvif = ifp;
1827 		mhead->m_pkthdr.csum_flags = 0;
1828 
1829 		/* Store the queue idx in the packet header. */
1830 		mhead->m_pkthdr.flowid = pq->kring_id;
1831 		M_HASHTYPE_SET(mhead, M_HASHTYPE_OPAQUE);
1832 
1833 		if (ifp->if_capenable & IFCAP_VLAN_HWTAGGING) {
1834 			struct ether_header *eh;
1835 
1836 			eh = mtod(mhead, struct ether_header *);
1837 			if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
1838 				ptnet_vlan_tag_remove(mhead);
1839 				/*
1840 				 * With the 802.1Q header removed, update the
1841 				 * checksum starting location accordingly.
1842 				 */
1843 				if (vh->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM)
1844 					vh->csum_start -= ETHER_VLAN_ENCAP_LEN;
1845 			}
1846 		}
1847 
1848 		if (unlikely(have_vnet_hdr && virtio_net_rx_csum(mhead, vh))) {
1849 			m_freem(mhead);
1850 			nm_prlim(1, "Csum offload error: dropping");
1851 			pq->stats.iqdrops ++;
1852 			deliver = 0;
1853 		}
1854 
1855 skip:
1856 		count ++;
1857 		if (++batch_count >= PTNET_RX_BATCH) {
1858 			/* Some packets have been (or will be) pushed to the network
1859 			 * stack. We need to update the CSB to tell the host about
1860 			 * the new ring->cur and ring->head (RX buffer refill). */
1861 			ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
1862 			batch_count = 0;
1863 		}
1864 
1865 		if (likely(deliver))  {
1866 			pq->stats.packets ++;
1867 			pq->stats.bytes += mhead->m_pkthdr.len;
1868 
1869 			PTNET_Q_UNLOCK(pq);
1870 			(*ifp->if_input)(ifp, mhead);
1871 			PTNET_Q_LOCK(pq);
1872 			/* The ring->head index (and related indices) are
1873 			 * updated under pq lock by ptnet_ring_update().
1874 			 * Since we dropped the lock to call if_input(), we
1875 			 * must reload ring->head and restart processing the
1876 			 * ring from there. */
1877 			head = ring->head;
1878 
1879 			if (unlikely(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) {
1880 				/* The interface has gone down while we didn't
1881 				 * have the lock. Stop any processing and exit. */
1882 				goto unlock;
1883 			}
1884 		}
1885 	}
1886 escape:
1887 	if (batch_count) {
1888 		ptnet_ring_update(pq, kring, head, NAF_FORCE_READ);
1889 
1890 	}
1891 
1892 	if (count >= budget && may_resched) {
1893 		/* If we ran out of budget or the double-check found new
1894 		 * slots to process, schedule the taskqueue. */
1895 		DBG(nm_prlim(1, "out of budget: resched h %u t %u\n",
1896 					head, ring->tail));
1897 		taskqueue_enqueue(pq->taskq, &pq->task);
1898 	}
1899 unlock:
1900 	PTNET_Q_UNLOCK(pq);
1901 
1902 	return count;
1903 }
1904 
1905 static void
ptnet_rx_task(void * context,int pending)1906 ptnet_rx_task(void *context, int pending)
1907 {
1908 	struct ptnet_queue *pq = context;
1909 
1910 	DBG(nm_prlim(1, "%s: pq #%u\n", __func__, pq->kring_id));
1911 	ptnet_rx_eof(pq, PTNET_RX_BUDGET, true);
1912 }
1913 
1914 static void
ptnet_tx_task(void * context,int pending)1915 ptnet_tx_task(void *context, int pending)
1916 {
1917 	struct ptnet_queue *pq = context;
1918 
1919 	DBG(nm_prlim(1, "%s: pq #%u\n", __func__, pq->kring_id));
1920 	ptnet_drain_transmit_queue(pq, PTNET_TX_BUDGET, true);
1921 }
1922 
1923 #ifdef DEVICE_POLLING
1924 /* We don't need to handle differently POLL_AND_CHECK_STATUS and
1925  * POLL_ONLY, since we don't have an Interrupt Status Register. */
1926 static int
ptnet_poll(if_t ifp,enum poll_cmd cmd,int budget)1927 ptnet_poll(if_t ifp, enum poll_cmd cmd, int budget)
1928 {
1929 	struct ptnet_softc *sc = if_getsoftc(ifp);
1930 	unsigned int queue_budget;
1931 	unsigned int count = 0;
1932 	bool borrow = false;
1933 	int i;
1934 
1935 	KASSERT(sc->num_rings > 0, ("Found no queues in while polling ptnet"));
1936 	queue_budget = MAX(budget / sc->num_rings, 1);
1937 	nm_prlim(1, "Per-queue budget is %d", queue_budget);
1938 
1939 	while (budget) {
1940 		unsigned int rcnt = 0;
1941 
1942 		for (i = 0; i < sc->num_rings; i++) {
1943 			struct ptnet_queue *pq = sc->queues + i;
1944 
1945 			if (borrow) {
1946 				queue_budget = MIN(queue_budget, budget);
1947 				if (queue_budget == 0) {
1948 					break;
1949 				}
1950 			}
1951 
1952 			if (i < sc->num_tx_rings) {
1953 				rcnt += ptnet_drain_transmit_queue(pq,
1954 						   queue_budget, false);
1955 			} else {
1956 				rcnt += ptnet_rx_eof(pq, queue_budget,
1957 						      false);
1958 			}
1959 		}
1960 
1961 		if (!rcnt) {
1962 			/* A scan of the queues gave no result, we can
1963 			 * stop here. */
1964 			break;
1965 		}
1966 
1967 		if (rcnt > budget) {
1968 			/* This may happen when initial budget < sc->num_rings,
1969 			 * since one packet budget is given to each queue
1970 			 * anyway. Just pretend we didn't eat "so much". */
1971 			rcnt = budget;
1972 		}
1973 		count += rcnt;
1974 		budget -= rcnt;
1975 		borrow = true;
1976 	}
1977 
1978 
1979 	return count;
1980 }
1981 #endif /* DEVICE_POLLING */
1982 #endif /* WITH_PTNETMAP */
1983