1 /*
2  * Copyright (C) 2014-2018 Vincenzo Maffione, Luigi Rizzo.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  */
25 
26 /*
27  * $FreeBSD$
28  */
29 
30 #include <net/netmap.h>
31 #include <sys/selinfo.h>
32 #include <vm/vm.h>
33 #include <vm/pmap.h>    /* vtophys ? */
34 #include <dev/netmap/netmap_kern.h>
35 
36 /*
37  * Return 1 if the queue identified by 't' and 'idx' is in netmap mode.
38  */
39 static int
40 vtnet_netmap_queue_on(struct vtnet_softc *sc, enum txrx t, int idx)
41 {
42 	struct netmap_adapter *na = NA(sc->vtnet_ifp);
43 
44 	if (!nm_native_on(na))
45 		return 0;
46 
47 	if (t == NR_RX)
48 		return !!(idx < na->num_rx_rings &&
49 			na->rx_rings[idx]->nr_mode == NKR_NETMAP_ON);
50 
51 	return !!(idx < na->num_tx_rings &&
52 		na->tx_rings[idx]->nr_mode == NKR_NETMAP_ON);
53 }
54 
55 /* Register and unregister. */
56 static int
57 vtnet_netmap_reg(struct netmap_adapter *na, int state)
58 {
59 	struct ifnet *ifp = na->ifp;
60 	struct vtnet_softc *sc = ifp->if_softc;
61 	int success;
62 	int i;
63 
64 	/* Drain the taskqueues to make sure that there are no worker threads
65 	 * accessing the virtqueues. */
66 	vtnet_drain_taskqueues(sc);
67 
68 	VTNET_CORE_LOCK(sc);
69 
70 	/* We need nm_netmap_on() to return true when called by
71 	 * vtnet_init_locked() below. */
72 	if (state)
73 		nm_set_native_flags(na);
74 
75 	/* We need to trigger a device reset in order to unexpose guest buffers
76 	 * published to the host. */
77 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
78 	/* Get pending used buffers. The way they are freed depends on whether
79 	 * they are netmap buffer or they are mbufs. We can tell apart the two
80 	 * cases by looking at kring->nr_mode, before this is possibly updated
81 	 * in the loop below. */
82 	for (i = 0; i < sc->vtnet_act_vq_pairs; i++) {
83 		struct vtnet_txq *txq = &sc->vtnet_txqs[i];
84 		struct vtnet_rxq *rxq = &sc->vtnet_rxqs[i];
85 
86 		VTNET_TXQ_LOCK(txq);
87 		vtnet_txq_free_mbufs(txq);
88 		VTNET_TXQ_UNLOCK(txq);
89 
90 		VTNET_RXQ_LOCK(rxq);
91 		vtnet_rxq_free_mbufs(rxq);
92 		VTNET_RXQ_UNLOCK(rxq);
93 	}
94 	vtnet_init_locked(sc);
95 	success = (ifp->if_drv_flags & IFF_DRV_RUNNING) ? 0 : ENXIO;
96 
97 	if (state) {
98 		netmap_krings_mode_commit(na, state);
99 	} else {
100 		nm_clear_native_flags(na);
101 		netmap_krings_mode_commit(na, state);
102 	}
103 
104 	VTNET_CORE_UNLOCK(sc);
105 
106 	return success;
107 }
108 
109 
110 /* Reconcile kernel and user view of the transmit ring. */
111 static int
112 vtnet_netmap_txsync(struct netmap_kring *kring, int flags)
113 {
114 	struct netmap_adapter *na = kring->na;
115 	struct ifnet *ifp = na->ifp;
116 	struct netmap_ring *ring = kring->ring;
117 	u_int ring_nr = kring->ring_id;
118 	u_int nm_i;	/* index into the netmap ring */
119 	u_int const lim = kring->nkr_num_slots - 1;
120 	u_int const head = kring->rhead;
121 
122 	/* device-specific */
123 	struct vtnet_softc *sc = ifp->if_softc;
124 	struct vtnet_txq *txq = &sc->vtnet_txqs[ring_nr];
125 	struct virtqueue *vq = txq->vtntx_vq;
126 	int interrupts = !(kring->nr_kflags & NKR_NOINTR);
127 	u_int n;
128 
129 	/*
130 	 * First part: process new packets to send.
131 	 */
132 	rmb();
133 
134 	nm_i = kring->nr_hwcur;
135 	if (nm_i != head) {	/* we have new packets to send */
136 		struct sglist *sg = txq->vtntx_sg;
137 
138 		for (; nm_i != head; nm_i = nm_next(nm_i, lim)) {
139 			/* we use an empty header here */
140 			struct netmap_slot *slot = &ring->slot[nm_i];
141 			u_int len = slot->len;
142 			uint64_t paddr;
143 			void *addr = PNMB(na, slot, &paddr);
144 			int err;
145 
146 			NM_CHECK_ADDR_LEN(na, addr, len);
147 
148 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
149 			/* Initialize the scatterlist, expose it to the hypervisor,
150 			 * and kick the hypervisor (if necessary).
151 			 */
152 			sglist_reset(sg); // cheap
153 			err = sglist_append(sg, &txq->vtntx_shrhdr, sc->vtnet_hdr_size);
154 			err |= sglist_append_phys(sg, paddr, len);
155 			KASSERT(err == 0, ("%s: cannot append to sglist %d",
156 						__func__, err));
157 			err = virtqueue_enqueue(vq, /*cookie=*/txq, sg,
158 						/*readable=*/sg->sg_nseg,
159 						/*writeable=*/0);
160 			if (unlikely(err)) {
161 				if (err != ENOSPC)
162 					nm_prerr("virtqueue_enqueue(%s) failed: %d",
163 							kring->name, err);
164 				break;
165 			}
166 		}
167 
168 		virtqueue_notify(vq);
169 
170 		/* Update hwcur depending on where we stopped. */
171 		kring->nr_hwcur = nm_i; /* note we migth break early */
172 	}
173 
174 	/* Free used slots. We only consider our own used buffers, recognized
175 	 * by the token we passed to virtqueue_enqueue.
176 	 */
177 	n = 0;
178 	for (;;) {
179 		void *token = virtqueue_dequeue(vq, NULL);
180 		if (token == NULL)
181 			break;
182 		if (unlikely(token != (void *)txq))
183 			nm_prerr("BUG: TX token mismatch");
184 		else
185 			n++;
186 	}
187 	if (n > 0) {
188 		kring->nr_hwtail += n;
189 		if (kring->nr_hwtail > lim)
190 			kring->nr_hwtail -= lim + 1;
191 	}
192 
193 	if (interrupts && virtqueue_nfree(vq) < 32)
194 		virtqueue_postpone_intr(vq, VQ_POSTPONE_LONG);
195 
196 	return 0;
197 }
198 
199 /*
200  * Publish (up to) num netmap receive buffers to the host,
201  * starting from the first one that the user made available
202  * (kring->nr_hwcur).
203  */
204 static int
205 vtnet_netmap_kring_refill(struct netmap_kring *kring, u_int num)
206 {
207 	struct netmap_adapter *na = kring->na;
208 	struct ifnet *ifp = na->ifp;
209 	struct netmap_ring *ring = kring->ring;
210 	u_int ring_nr = kring->ring_id;
211 	u_int const lim = kring->nkr_num_slots - 1;
212 	u_int nm_i = kring->nr_hwcur;
213 
214 	/* device-specific */
215 	struct vtnet_softc *sc = ifp->if_softc;
216 	struct vtnet_rxq *rxq = &sc->vtnet_rxqs[ring_nr];
217 	struct virtqueue *vq = rxq->vtnrx_vq;
218 
219 	/* use a local sglist, default might be short */
220 	struct sglist_seg ss[2];
221 	struct sglist sg = { ss, 0, 0, 2 };
222 
223 	for (; num > 0; nm_i = nm_next(nm_i, lim), num--) {
224 		struct netmap_slot *slot = &ring->slot[nm_i];
225 		uint64_t paddr;
226 		void *addr = PNMB(na, slot, &paddr);
227 		int err;
228 
229 		if (addr == NETMAP_BUF_BASE(na)) { /* bad buf */
230 			if (netmap_ring_reinit(kring))
231 				return -1;
232 		}
233 
234 		slot->flags &= ~NS_BUF_CHANGED;
235 		sglist_reset(&sg);
236 		err = sglist_append(&sg, &rxq->vtnrx_shrhdr, sc->vtnet_hdr_size);
237 		err |= sglist_append_phys(&sg, paddr, NETMAP_BUF_SIZE(na));
238 		KASSERT(err == 0, ("%s: cannot append to sglist %d",
239 					__func__, err));
240 		/* writable for the host */
241 		err = virtqueue_enqueue(vq, /*cookie=*/rxq, &sg,
242 				/*readable=*/0, /*writeable=*/sg.sg_nseg);
243 		if (unlikely(err)) {
244 			if (err != ENOSPC)
245 				nm_prerr("virtqueue_enqueue(%s) failed: %d",
246 					kring->name, err);
247 			break;
248 		}
249 	}
250 
251 	return nm_i;
252 }
253 
254 /*
255  * Publish netmap buffers on a RX virtqueue.
256  * Returns -1 if this virtqueue is not being opened in netmap mode.
257  * If the virtqueue is being opened in netmap mode, return 0 on success and
258  * a positive error code on failure.
259  */
260 static int
261 vtnet_netmap_rxq_populate(struct vtnet_rxq *rxq)
262 {
263 	struct netmap_adapter *na = NA(rxq->vtnrx_sc->vtnet_ifp);
264 	struct netmap_kring *kring;
265 	int error;
266 
267 	if (!nm_native_on(na) || rxq->vtnrx_id >= na->num_rx_rings)
268 		return -1;
269 
270 	kring = na->rx_rings[rxq->vtnrx_id];
271 	if (!(nm_kring_pending_on(kring) ||
272 			kring->nr_pending_mode == NKR_NETMAP_ON))
273 		return -1;
274 
275 	/* Expose all the RX netmap buffers we can. In case of no indirect
276 	 * buffers, the number of netmap slots in the RX ring matches the
277 	 * maximum number of 2-elements sglist that the RX virtqueue can
278 	 * accommodate. */
279 	error = vtnet_netmap_kring_refill(kring, na->num_rx_desc);
280 	virtqueue_notify(rxq->vtnrx_vq);
281 
282 	return error < 0 ? ENXIO : 0;
283 }
284 
285 /* Reconcile kernel and user view of the receive ring. */
286 static int
287 vtnet_netmap_rxsync(struct netmap_kring *kring, int flags)
288 {
289 	struct netmap_adapter *na = kring->na;
290 	struct ifnet *ifp = na->ifp;
291 	struct netmap_ring *ring = kring->ring;
292 	u_int ring_nr = kring->ring_id;
293 	u_int nm_i;	/* index into the netmap ring */
294 	u_int const lim = kring->nkr_num_slots - 1;
295 	u_int const head = kring->rhead;
296 	int force_update = (flags & NAF_FORCE_READ) ||
297 				(kring->nr_kflags & NKR_PENDINTR);
298 	int interrupts = !(kring->nr_kflags & NKR_NOINTR);
299 
300 	/* device-specific */
301 	struct vtnet_softc *sc = ifp->if_softc;
302 	struct vtnet_rxq *rxq = &sc->vtnet_rxqs[ring_nr];
303 	struct virtqueue *vq = rxq->vtnrx_vq;
304 
305 	rmb();
306 	/*
307 	 * First part: import newly received packets.
308 	 * Only accept our own buffers (matching the token). We should only get
309 	 * matching buffers. We may need to stop early to avoid hwtail to overrun
310 	 * hwcur.
311 	 */
312 	if (netmap_no_pendintr || force_update) {
313 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
314 		void *token;
315 
316 		vtnet_rxq_disable_intr(rxq);
317 
318 		nm_i = kring->nr_hwtail;
319 		while (nm_i != hwtail_lim) {
320 			int len;
321 			token = virtqueue_dequeue(vq, &len);
322 			if (token == NULL) {
323 				if (interrupts && vtnet_rxq_enable_intr(rxq)) {
324 					vtnet_rxq_disable_intr(rxq);
325 					continue;
326 				}
327 				break;
328 			}
329 			if (unlikely(token != (void *)rxq)) {
330 				nm_prerr("BUG: RX token mismatch");
331 			} else {
332 				/* Skip the virtio-net header. */
333 				len -= sc->vtnet_hdr_size;
334 				if (unlikely(len < 0)) {
335 					nm_prlim(1, "Truncated virtio-net-header, "
336 						"missing %d bytes", -len);
337 					len = 0;
338 				}
339 				ring->slot[nm_i].len = len;
340 				ring->slot[nm_i].flags = 0;
341 				nm_i = nm_next(nm_i, lim);
342 			}
343 		}
344 		kring->nr_hwtail = nm_i;
345 		kring->nr_kflags &= ~NKR_PENDINTR;
346 	}
347 	nm_prdis("[B] h %d c %d hwcur %d hwtail %d", ring->head, ring->cur,
348 				kring->nr_hwcur, kring->nr_hwtail);
349 
350 	/*
351 	 * Second part: skip past packets that userspace has released.
352 	 */
353 	nm_i = kring->nr_hwcur; /* netmap ring index */
354 	if (nm_i != head) {
355 		int howmany = head - nm_i;
356 		int nm_j;
357 
358 		if (howmany < 0)
359 			howmany += kring->nkr_num_slots;
360 		nm_j = vtnet_netmap_kring_refill(kring, howmany);
361 		if (nm_j < 0)
362 			return nm_j;
363 		kring->nr_hwcur = nm_j;
364 		virtqueue_notify(vq);
365 	}
366 
367 	nm_prdis("[C] h %d c %d t %d hwcur %d hwtail %d", ring->head, ring->cur,
368 		ring->tail, kring->nr_hwcur, kring->nr_hwtail);
369 
370 	return 0;
371 }
372 
373 
374 /* Enable/disable interrupts on all virtqueues. */
375 static void
376 vtnet_netmap_intr(struct netmap_adapter *na, int state)
377 {
378 	struct vtnet_softc *sc = na->ifp->if_softc;
379 	int i;
380 
381 	for (i = 0; i < sc->vtnet_max_vq_pairs; i++) {
382 		struct vtnet_rxq *rxq = &sc->vtnet_rxqs[i];
383 		struct vtnet_txq *txq = &sc->vtnet_txqs[i];
384 		struct virtqueue *txvq = txq->vtntx_vq;
385 
386 		if (state) {
387 			vtnet_rxq_enable_intr(rxq);
388 			virtqueue_enable_intr(txvq);
389 		} else {
390 			vtnet_rxq_disable_intr(rxq);
391 			virtqueue_disable_intr(txvq);
392 		}
393 	}
394 }
395 
396 static int
397 vtnet_netmap_tx_slots(struct vtnet_softc *sc)
398 {
399 	int div;
400 
401 	/* We need to prepend a virtio-net header to each netmap buffer to be
402 	 * transmitted, therefore calling virtqueue_enqueue() passing sglist
403 	 * with 2 elements.
404 	 * TX virtqueues use indirect descriptors if the feature was negotiated
405 	 * with the host, and if sc->vtnet_tx_nsegs > 1. With indirect
406 	 * descriptors, a single virtio descriptor is sufficient to reference
407 	 * each TX sglist. Without them, we need two separate virtio descriptors
408 	 * for each TX sglist. We therefore compute the number of netmap TX
409 	 * slots according to these assumptions.
410 	 */
411 	if ((sc->vtnet_flags & VTNET_FLAG_INDIRECT) && sc->vtnet_tx_nsegs > 1)
412 		div = 1;
413 	else
414 		div = 2;
415 
416 	return virtqueue_size(sc->vtnet_txqs[0].vtntx_vq) / div;
417 }
418 
419 static int
420 vtnet_netmap_rx_slots(struct vtnet_softc *sc)
421 {
422 	int div;
423 
424 	/* We need to prepend a virtio-net header to each netmap buffer to be
425 	 * received, therefore calling virtqueue_enqueue() passing sglist
426 	 * with 2 elements.
427 	 * RX virtqueues use indirect descriptors if the feature was negotiated
428 	 * with the host, and if sc->vtnet_rx_nsegs > 1. With indirect
429 	 * descriptors, a single virtio descriptor is sufficient to reference
430 	 * each RX sglist. Without them, we need two separate virtio descriptors
431 	 * for each RX sglist. We therefore compute the number of netmap RX
432 	 * slots according to these assumptions.
433 	 */
434 	if ((sc->vtnet_flags & VTNET_FLAG_INDIRECT) && sc->vtnet_rx_nsegs > 1)
435 		div = 1;
436 	else
437 		div = 2;
438 
439 	return virtqueue_size(sc->vtnet_rxqs[0].vtnrx_vq) / div;
440 }
441 
442 static int
443 vtnet_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
444 {
445 	struct vtnet_softc *sc = na->ifp->if_softc;
446 
447 	info->num_tx_rings = sc->vtnet_act_vq_pairs;
448 	info->num_rx_rings = sc->vtnet_act_vq_pairs;
449 	info->num_tx_descs = vtnet_netmap_tx_slots(sc);
450 	info->num_rx_descs = vtnet_netmap_rx_slots(sc);
451 	info->rx_buf_maxsize = NETMAP_BUF_SIZE(na);
452 
453 	return 0;
454 }
455 
456 static void
457 vtnet_netmap_attach(struct vtnet_softc *sc)
458 {
459 	struct netmap_adapter na;
460 
461 	bzero(&na, sizeof(na));
462 
463 	na.ifp = sc->vtnet_ifp;
464 	na.na_flags = 0;
465 	na.num_tx_desc = vtnet_netmap_tx_slots(sc);
466 	na.num_rx_desc = vtnet_netmap_rx_slots(sc);
467 	na.num_tx_rings = na.num_rx_rings = sc->vtnet_max_vq_pairs;
468 	na.rx_buf_maxsize = 0;
469 	na.nm_register = vtnet_netmap_reg;
470 	na.nm_txsync = vtnet_netmap_txsync;
471 	na.nm_rxsync = vtnet_netmap_rxsync;
472 	na.nm_intr = vtnet_netmap_intr;
473 	na.nm_config = vtnet_netmap_config;
474 
475 	netmap_attach(&na);
476 
477 	nm_prinf("vtnet attached txq=%d, txd=%d rxq=%d, rxd=%d",
478 			na.num_tx_rings, na.num_tx_desc,
479 			na.num_tx_rings, na.num_rx_desc);
480 }
481 /* end of file */
482