1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2004-2006 Kip Macy
5 * Copyright (c) 2015 Wei Liu <wei.liu2@citrix.com>
6 * All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35
36 #include <sys/param.h>
37 #include <sys/sockio.h>
38 #include <sys/limits.h>
39 #include <sys/mbuf.h>
40 #include <sys/malloc.h>
41 #include <sys/module.h>
42 #include <sys/kernel.h>
43 #include <sys/socket.h>
44 #include <sys/sysctl.h>
45 #include <sys/taskqueue.h>
46
47 #include <net/if.h>
48 #include <net/if_var.h>
49 #include <net/if_arp.h>
50 #include <net/ethernet.h>
51 #include <net/if_media.h>
52 #include <net/bpf.h>
53 #include <net/if_types.h>
54
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/if_ether.h>
58 #include <netinet/tcp.h>
59 #include <netinet/tcp_lro.h>
60
61 #include <vm/vm.h>
62 #include <vm/pmap.h>
63
64 #include <sys/bus.h>
65
66 #include <xen/xen-os.h>
67 #include <xen/hypervisor.h>
68 #include <xen/xen_intr.h>
69 #include <xen/gnttab.h>
70 #include <contrib/xen/memory.h>
71 #include <contrib/xen/io/netif.h>
72 #include <xen/xenbus/xenbusvar.h>
73
74 #include <machine/bus.h>
75
76 #include "xenbus_if.h"
77
78 /* Features supported by all backends. TSO and LRO can be negotiated */
79 #define XN_CSUM_FEATURES (CSUM_TCP | CSUM_UDP)
80
81 #define NET_TX_RING_SIZE __CONST_RING_SIZE(netif_tx, PAGE_SIZE)
82 #define NET_RX_RING_SIZE __CONST_RING_SIZE(netif_rx, PAGE_SIZE)
83
84 #define NET_RX_SLOTS_MIN (XEN_NETIF_NR_SLOTS_MIN + 1)
85
86 /*
87 * Should the driver do LRO on the RX end
88 * this can be toggled on the fly, but the
89 * interface must be reset (down/up) for it
90 * to take effect.
91 */
92 static int xn_enable_lro = 1;
93 TUNABLE_INT("hw.xn.enable_lro", &xn_enable_lro);
94
95 /*
96 * Number of pairs of queues.
97 */
98 static unsigned long xn_num_queues = 4;
99 TUNABLE_ULONG("hw.xn.num_queues", &xn_num_queues);
100
101 /**
102 * \brief The maximum allowed data fragments in a single transmit
103 * request.
104 *
105 * This limit is imposed by the backend driver. We assume here that
106 * we are dealing with a Linux driver domain and have set our limit
107 * to mirror the Linux MAX_SKB_FRAGS constant.
108 */
109 #define MAX_TX_REQ_FRAGS (65536 / PAGE_SIZE + 2)
110
111 #define RX_COPY_THRESHOLD 256
112
113 #define net_ratelimit() 0
114
115 struct netfront_rxq;
116 struct netfront_txq;
117 struct netfront_info;
118 struct netfront_rx_info;
119
120 static void xn_txeof(struct netfront_txq *);
121 static void xn_rxeof(struct netfront_rxq *);
122 static void xn_alloc_rx_buffers(struct netfront_rxq *);
123 static void xn_alloc_rx_buffers_callout(void *arg);
124
125 static void xn_release_rx_bufs(struct netfront_rxq *);
126 static void xn_release_tx_bufs(struct netfront_txq *);
127
128 static void xn_rxq_intr(struct netfront_rxq *);
129 static void xn_txq_intr(struct netfront_txq *);
130 static void xn_intr(void *);
131 static int xn_assemble_tx_request(struct netfront_txq *, struct mbuf *);
132 static int xn_ioctl(struct ifnet *, u_long, caddr_t);
133 static void xn_ifinit_locked(struct netfront_info *);
134 static void xn_ifinit(void *);
135 static void xn_stop(struct netfront_info *);
136 static void xn_query_features(struct netfront_info *np);
137 static int xn_configure_features(struct netfront_info *np);
138 static void netif_free(struct netfront_info *info);
139 static int netfront_detach(device_t dev);
140
141 static int xn_txq_mq_start_locked(struct netfront_txq *, struct mbuf *);
142 static int xn_txq_mq_start(struct ifnet *, struct mbuf *);
143
144 static int talk_to_backend(device_t dev, struct netfront_info *info);
145 static int create_netdev(device_t dev);
146 static void netif_disconnect_backend(struct netfront_info *info);
147 static int setup_device(device_t dev, struct netfront_info *info,
148 unsigned long);
149 static int xn_ifmedia_upd(struct ifnet *ifp);
150 static void xn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr);
151
152 static int xn_connect(struct netfront_info *);
153 static void xn_kick_rings(struct netfront_info *);
154
155 static int xn_get_responses(struct netfront_rxq *,
156 struct netfront_rx_info *, RING_IDX, RING_IDX *,
157 struct mbuf **);
158
159 #define virt_to_mfn(x) (vtophys(x) >> PAGE_SHIFT)
160
161 #define INVALID_P2M_ENTRY (~0UL)
162 #define XN_QUEUE_NAME_LEN 8 /* xn{t,r}x_%u, allow for two digits */
163 struct netfront_rxq {
164 struct netfront_info *info;
165 u_int id;
166 char name[XN_QUEUE_NAME_LEN];
167 struct mtx lock;
168
169 int ring_ref;
170 netif_rx_front_ring_t ring;
171 xen_intr_handle_t xen_intr_handle;
172
173 grant_ref_t gref_head;
174 grant_ref_t grant_ref[NET_RX_RING_SIZE + 1];
175
176 struct mbuf *mbufs[NET_RX_RING_SIZE + 1];
177
178 struct lro_ctrl lro;
179
180 struct callout rx_refill;
181 };
182
183 struct netfront_txq {
184 struct netfront_info *info;
185 u_int id;
186 char name[XN_QUEUE_NAME_LEN];
187 struct mtx lock;
188
189 int ring_ref;
190 netif_tx_front_ring_t ring;
191 xen_intr_handle_t xen_intr_handle;
192
193 grant_ref_t gref_head;
194 grant_ref_t grant_ref[NET_TX_RING_SIZE + 1];
195
196 struct mbuf *mbufs[NET_TX_RING_SIZE + 1];
197 int mbufs_cnt;
198 struct buf_ring *br;
199
200 struct taskqueue *tq;
201 struct task defrtask;
202
203 bus_dma_segment_t segs[MAX_TX_REQ_FRAGS];
204 struct mbuf_xennet {
205 struct m_tag tag;
206 bus_dma_tag_t dma_tag;
207 bus_dmamap_t dma_map;
208 struct netfront_txq *txq;
209 SLIST_ENTRY(mbuf_xennet) next;
210 u_int count;
211 } xennet_tag[NET_TX_RING_SIZE + 1];
212 SLIST_HEAD(, mbuf_xennet) tags;
213
214 bool full;
215 };
216
217 struct netfront_info {
218 struct ifnet *xn_ifp;
219
220 struct mtx sc_lock;
221
222 u_int num_queues;
223 struct netfront_rxq *rxq;
224 struct netfront_txq *txq;
225
226 u_int carrier;
227 u_int maxfrags;
228
229 device_t xbdev;
230 uint8_t mac[ETHER_ADDR_LEN];
231
232 int xn_if_flags;
233
234 struct ifmedia sc_media;
235
236 bus_dma_tag_t dma_tag;
237
238 bool xn_reset;
239 };
240
241 struct netfront_rx_info {
242 struct netif_rx_response rx;
243 struct netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
244 };
245
246 #define XN_RX_LOCK(_q) mtx_lock(&(_q)->lock)
247 #define XN_RX_UNLOCK(_q) mtx_unlock(&(_q)->lock)
248
249 #define XN_TX_LOCK(_q) mtx_lock(&(_q)->lock)
250 #define XN_TX_TRYLOCK(_q) mtx_trylock(&(_q)->lock)
251 #define XN_TX_UNLOCK(_q) mtx_unlock(&(_q)->lock)
252
253 #define XN_LOCK(_sc) mtx_lock(&(_sc)->sc_lock);
254 #define XN_UNLOCK(_sc) mtx_unlock(&(_sc)->sc_lock);
255
256 #define XN_LOCK_ASSERT(_sc) mtx_assert(&(_sc)->sc_lock, MA_OWNED);
257 #define XN_RX_LOCK_ASSERT(_q) mtx_assert(&(_q)->lock, MA_OWNED);
258 #define XN_TX_LOCK_ASSERT(_q) mtx_assert(&(_q)->lock, MA_OWNED);
259
260 #define netfront_carrier_on(netif) ((netif)->carrier = 1)
261 #define netfront_carrier_off(netif) ((netif)->carrier = 0)
262 #define netfront_carrier_ok(netif) ((netif)->carrier)
263
264 /* Access macros for acquiring freeing slots in xn_free_{tx,rx}_idxs[]. */
265
266 static inline void
267 add_id_to_freelist(struct mbuf **list, uintptr_t id)
268 {
269
270 KASSERT(id != 0,
271 ("%s: the head item (0) must always be free.", __func__));
272 list[id] = list[0];
273 list[0] = (struct mbuf *)id;
274 }
275
276 static inline unsigned short
277 get_id_from_freelist(struct mbuf **list)
278 {
279 uintptr_t id;
280
281 id = (uintptr_t)list[0];
282 KASSERT(id != 0,
283 ("%s: the head item (0) must always remain free.", __func__));
284 list[0] = list[id];
285 return (id);
286 }
287
288 static inline int
289 xn_rxidx(RING_IDX idx)
290 {
291
292 return idx & (NET_RX_RING_SIZE - 1);
293 }
294
295 static inline struct mbuf *
296 xn_get_rx_mbuf(struct netfront_rxq *rxq, RING_IDX ri)
297 {
298 int i;
299 struct mbuf *m;
300
301 i = xn_rxidx(ri);
302 m = rxq->mbufs[i];
303 rxq->mbufs[i] = NULL;
304 return (m);
305 }
306
307 static inline grant_ref_t
308 xn_get_rx_ref(struct netfront_rxq *rxq, RING_IDX ri)
309 {
310 int i = xn_rxidx(ri);
311 grant_ref_t ref = rxq->grant_ref[i];
312
313 KASSERT(ref != GRANT_REF_INVALID, ("Invalid grant reference!\n"));
314 rxq->grant_ref[i] = GRANT_REF_INVALID;
315 return (ref);
316 }
317
318 #define MTAG_COOKIE 1218492000
319 #define MTAG_XENNET 0
320
321 static void mbuf_grab(struct mbuf *m)
322 {
323 struct mbuf_xennet *ref;
324
325 ref = (struct mbuf_xennet *)m_tag_locate(m, MTAG_COOKIE,
326 MTAG_XENNET, NULL);
327 KASSERT(ref != NULL, ("Cannot find refcount"));
328 ref->count++;
329 }
330
331 static void mbuf_release(struct mbuf *m)
332 {
333 struct mbuf_xennet *ref;
334
335 ref = (struct mbuf_xennet *)m_tag_locate(m, MTAG_COOKIE,
336 MTAG_XENNET, NULL);
337 KASSERT(ref != NULL, ("Cannot find refcount"));
338 KASSERT(ref->count > 0, ("Invalid reference count"));
339
340 if (--ref->count == 0)
341 m_freem(m);
342 }
343
344 static void tag_free(struct m_tag *t)
345 {
346 struct mbuf_xennet *ref = (struct mbuf_xennet *)t;
347
348 KASSERT(ref->count == 0, ("Free mbuf tag with pending refcnt"));
349 bus_dmamap_sync(ref->dma_tag, ref->dma_map, BUS_DMASYNC_POSTWRITE);
350 bus_dmamap_destroy(ref->dma_tag, ref->dma_map);
351 SLIST_INSERT_HEAD(&ref->txq->tags, ref, next);
352 }
353
354 #define IPRINTK(fmt, args...) \
355 printf("[XEN] " fmt, ##args)
356 #ifdef INVARIANTS
357 #define WPRINTK(fmt, args...) \
358 printf("[XEN] " fmt, ##args)
359 #else
360 #define WPRINTK(fmt, args...)
361 #endif
362 #ifdef DEBUG
363 #define DPRINTK(fmt, args...) \
364 printf("[XEN] %s: " fmt, __func__, ##args)
365 #else
366 #define DPRINTK(fmt, args...)
367 #endif
368
369 /**
370 * Read the 'mac' node at the given device's node in the store, and parse that
371 * as colon-separated octets, placing result the given mac array. mac must be
372 * a preallocated array of length ETH_ALEN (as declared in linux/if_ether.h).
373 * Return 0 on success, or errno on error.
374 */
375 static int
376 xen_net_read_mac(device_t dev, uint8_t mac[])
377 {
378 int error, i;
379 char *s, *e, *macstr;
380 const char *path;
381
382 path = xenbus_get_node(dev);
383 error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
384 if (error == ENOENT) {
385 /*
386 * Deal with missing mac XenStore nodes on devices with
387 * HVM emulation (the 'ioemu' configuration attribute)
388 * enabled.
389 *
390 * The HVM emulator may execute in a stub device model
391 * domain which lacks the permission, only given to Dom0,
392 * to update the guest's XenStore tree. For this reason,
393 * the HVM emulator doesn't even attempt to write the
394 * front-side mac node, even when operating in Dom0.
395 * However, there should always be a mac listed in the
396 * backend tree. Fallback to this version if our query
397 * of the front side XenStore location doesn't find
398 * anything.
399 */
400 path = xenbus_get_otherend_path(dev);
401 error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
402 }
403 if (error != 0) {
404 xenbus_dev_fatal(dev, error, "parsing %s/mac", path);
405 return (error);
406 }
407
408 s = macstr;
409 for (i = 0; i < ETHER_ADDR_LEN; i++) {
410 mac[i] = strtoul(s, &e, 16);
411 if (s == e || (e[0] != ':' && e[0] != 0)) {
412 free(macstr, M_XENBUS);
413 return (ENOENT);
414 }
415 s = &e[1];
416 }
417 free(macstr, M_XENBUS);
418 return (0);
419 }
420
421 /**
422 * Entry point to this code when a new device is created. Allocate the basic
423 * structures and the ring buffers for communication with the backend, and
424 * inform the backend of the appropriate details for those. Switch to
425 * Connected state.
426 */
427 static int
428 netfront_probe(device_t dev)
429 {
430
431 if (xen_pv_nics_disabled())
432 return (ENXIO);
433
434 if (!strcmp(xenbus_get_type(dev), "vif")) {
435 device_set_desc(dev, "Virtual Network Interface");
436 return (0);
437 }
438
439 return (ENXIO);
440 }
441
442 static int
443 netfront_attach(device_t dev)
444 {
445 int err;
446
447 err = create_netdev(dev);
448 if (err != 0) {
449 xenbus_dev_fatal(dev, err, "creating netdev");
450 return (err);
451 }
452
453 SYSCTL_ADD_INT(device_get_sysctl_ctx(dev),
454 SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
455 OID_AUTO, "enable_lro", CTLFLAG_RW,
456 &xn_enable_lro, 0, "Large Receive Offload");
457
458 SYSCTL_ADD_ULONG(device_get_sysctl_ctx(dev),
459 SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
460 OID_AUTO, "num_queues", CTLFLAG_RD,
461 &xn_num_queues, "Number of pairs of queues");
462
463 return (0);
464 }
465
466 static int
467 netfront_suspend(device_t dev)
468 {
469 struct netfront_info *np = device_get_softc(dev);
470 u_int i;
471
472 for (i = 0; i < np->num_queues; i++) {
473 XN_RX_LOCK(&np->rxq[i]);
474 XN_TX_LOCK(&np->txq[i]);
475 }
476 netfront_carrier_off(np);
477 for (i = 0; i < np->num_queues; i++) {
478 XN_RX_UNLOCK(&np->rxq[i]);
479 XN_TX_UNLOCK(&np->txq[i]);
480 }
481 return (0);
482 }
483
484 /**
485 * We are reconnecting to the backend, due to a suspend/resume, or a backend
486 * driver restart. We tear down our netif structure and recreate it, but
487 * leave the device-layer structures intact so that this is transparent to the
488 * rest of the kernel.
489 */
490 static int
491 netfront_resume(device_t dev)
492 {
493 struct netfront_info *info = device_get_softc(dev);
494 u_int i;
495
496 if (xen_suspend_cancelled) {
497 for (i = 0; i < info->num_queues; i++) {
498 XN_RX_LOCK(&info->rxq[i]);
499 XN_TX_LOCK(&info->txq[i]);
500 }
501 netfront_carrier_on(info);
502 for (i = 0; i < info->num_queues; i++) {
503 XN_RX_UNLOCK(&info->rxq[i]);
504 XN_TX_UNLOCK(&info->txq[i]);
505 }
506 return (0);
507 }
508
509 netif_disconnect_backend(info);
510 return (0);
511 }
512
513 static int
514 write_queue_xenstore_keys(device_t dev,
515 struct netfront_rxq *rxq,
516 struct netfront_txq *txq,
517 struct xs_transaction *xst, bool hierarchy)
518 {
519 int err;
520 const char *message;
521 const char *node = xenbus_get_node(dev);
522 char *path;
523 size_t path_size;
524
525 KASSERT(rxq->id == txq->id, ("Mismatch between RX and TX queue ids"));
526 /* Split event channel support is not yet there. */
527 KASSERT(rxq->xen_intr_handle == txq->xen_intr_handle,
528 ("Split event channels are not supported"));
529
530 if (hierarchy) {
531 path_size = strlen(node) + 10;
532 path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
533 snprintf(path, path_size, "%s/queue-%u", node, rxq->id);
534 } else {
535 path_size = strlen(node) + 1;
536 path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
537 snprintf(path, path_size, "%s", node);
538 }
539
540 err = xs_printf(*xst, path, "tx-ring-ref","%u", txq->ring_ref);
541 if (err != 0) {
542 message = "writing tx ring-ref";
543 goto error;
544 }
545 err = xs_printf(*xst, path, "rx-ring-ref","%u", rxq->ring_ref);
546 if (err != 0) {
547 message = "writing rx ring-ref";
548 goto error;
549 }
550 err = xs_printf(*xst, path, "event-channel", "%u",
551 xen_intr_port(rxq->xen_intr_handle));
552 if (err != 0) {
553 message = "writing event-channel";
554 goto error;
555 }
556
557 free(path, M_DEVBUF);
558
559 return (0);
560
561 error:
562 free(path, M_DEVBUF);
563 xenbus_dev_fatal(dev, err, "%s", message);
564
565 return (err);
566 }
567
568 /* Common code used when first setting up, and when resuming. */
569 static int
570 talk_to_backend(device_t dev, struct netfront_info *info)
571 {
572 const char *message;
573 struct xs_transaction xst;
574 const char *node = xenbus_get_node(dev);
575 int err;
576 unsigned long num_queues, max_queues = 0;
577 unsigned int i;
578
579 err = xen_net_read_mac(dev, info->mac);
580 if (err != 0) {
581 xenbus_dev_fatal(dev, err, "parsing %s/mac", node);
582 goto out;
583 }
584
585 err = xs_scanf(XST_NIL, xenbus_get_otherend_path(info->xbdev),
586 "multi-queue-max-queues", NULL, "%lu", &max_queues);
587 if (err != 0)
588 max_queues = 1;
589 num_queues = xn_num_queues;
590 if (num_queues > max_queues)
591 num_queues = max_queues;
592
593 err = setup_device(dev, info, num_queues);
594 if (err != 0)
595 goto out;
596
597 again:
598 err = xs_transaction_start(&xst);
599 if (err != 0) {
600 xenbus_dev_fatal(dev, err, "starting transaction");
601 goto free;
602 }
603
604 if (info->num_queues == 1) {
605 err = write_queue_xenstore_keys(dev, &info->rxq[0],
606 &info->txq[0], &xst, false);
607 if (err != 0)
608 goto abort_transaction_no_def_error;
609 } else {
610 err = xs_printf(xst, node, "multi-queue-num-queues",
611 "%u", info->num_queues);
612 if (err != 0) {
613 message = "writing multi-queue-num-queues";
614 goto abort_transaction;
615 }
616
617 for (i = 0; i < info->num_queues; i++) {
618 err = write_queue_xenstore_keys(dev, &info->rxq[i],
619 &info->txq[i], &xst, true);
620 if (err != 0)
621 goto abort_transaction_no_def_error;
622 }
623 }
624
625 err = xs_printf(xst, node, "request-rx-copy", "%u", 1);
626 if (err != 0) {
627 message = "writing request-rx-copy";
628 goto abort_transaction;
629 }
630 err = xs_printf(xst, node, "feature-rx-notify", "%d", 1);
631 if (err != 0) {
632 message = "writing feature-rx-notify";
633 goto abort_transaction;
634 }
635 err = xs_printf(xst, node, "feature-sg", "%d", 1);
636 if (err != 0) {
637 message = "writing feature-sg";
638 goto abort_transaction;
639 }
640 if ((info->xn_ifp->if_capenable & IFCAP_LRO) != 0) {
641 err = xs_printf(xst, node, "feature-gso-tcpv4", "%d", 1);
642 if (err != 0) {
643 message = "writing feature-gso-tcpv4";
644 goto abort_transaction;
645 }
646 }
647 if ((info->xn_ifp->if_capenable & IFCAP_RXCSUM) == 0) {
648 err = xs_printf(xst, node, "feature-no-csum-offload", "%d", 1);
649 if (err != 0) {
650 message = "writing feature-no-csum-offload";
651 goto abort_transaction;
652 }
653 }
654
655 err = xs_transaction_end(xst, 0);
656 if (err != 0) {
657 if (err == EAGAIN)
658 goto again;
659 xenbus_dev_fatal(dev, err, "completing transaction");
660 goto free;
661 }
662
663 return 0;
664
665 abort_transaction:
666 xenbus_dev_fatal(dev, err, "%s", message);
667 abort_transaction_no_def_error:
668 xs_transaction_end(xst, 1);
669 free:
670 netif_free(info);
671 out:
672 return (err);
673 }
674
675 static void
676 xn_rxq_intr(struct netfront_rxq *rxq)
677 {
678
679 XN_RX_LOCK(rxq);
680 xn_rxeof(rxq);
681 XN_RX_UNLOCK(rxq);
682 }
683
684 static void
685 xn_txq_start(struct netfront_txq *txq)
686 {
687 struct netfront_info *np = txq->info;
688 struct ifnet *ifp = np->xn_ifp;
689
690 XN_TX_LOCK_ASSERT(txq);
691 if (!drbr_empty(ifp, txq->br))
692 xn_txq_mq_start_locked(txq, NULL);
693 }
694
695 static void
696 xn_txq_intr(struct netfront_txq *txq)
697 {
698
699 XN_TX_LOCK(txq);
700 if (RING_HAS_UNCONSUMED_RESPONSES(&txq->ring))
701 xn_txeof(txq);
702 xn_txq_start(txq);
703 XN_TX_UNLOCK(txq);
704 }
705
706 static void
707 xn_txq_tq_deferred(void *xtxq, int pending)
708 {
709 struct netfront_txq *txq = xtxq;
710
711 XN_TX_LOCK(txq);
712 xn_txq_start(txq);
713 XN_TX_UNLOCK(txq);
714 }
715
716 static void
717 disconnect_rxq(struct netfront_rxq *rxq)
718 {
719
720 xn_release_rx_bufs(rxq);
721 gnttab_free_grant_references(rxq->gref_head);
722 gnttab_end_foreign_access(rxq->ring_ref, NULL);
723 /*
724 * No split event channel support at the moment, handle will
725 * be unbound in tx. So no need to call xen_intr_unbind here,
726 * but we do want to reset the handler to 0.
727 */
728 rxq->xen_intr_handle = 0;
729 }
730
731 static void
732 destroy_rxq(struct netfront_rxq *rxq)
733 {
734
735 callout_drain(&rxq->rx_refill);
736 free(rxq->ring.sring, M_DEVBUF);
737 }
738
739 static void
740 destroy_rxqs(struct netfront_info *np)
741 {
742 int i;
743
744 for (i = 0; i < np->num_queues; i++)
745 destroy_rxq(&np->rxq[i]);
746
747 free(np->rxq, M_DEVBUF);
748 np->rxq = NULL;
749 }
750
751 static int
752 setup_rxqs(device_t dev, struct netfront_info *info,
753 unsigned long num_queues)
754 {
755 int q, i;
756 int error;
757 netif_rx_sring_t *rxs;
758 struct netfront_rxq *rxq;
759
760 info->rxq = malloc(sizeof(struct netfront_rxq) * num_queues,
761 M_DEVBUF, M_WAITOK|M_ZERO);
762
763 for (q = 0; q < num_queues; q++) {
764 rxq = &info->rxq[q];
765
766 rxq->id = q;
767 rxq->info = info;
768 rxq->ring_ref = GRANT_REF_INVALID;
769 rxq->ring.sring = NULL;
770 snprintf(rxq->name, XN_QUEUE_NAME_LEN, "xnrx_%u", q);
771 mtx_init(&rxq->lock, rxq->name, "netfront receive lock",
772 MTX_DEF);
773
774 for (i = 0; i <= NET_RX_RING_SIZE; i++) {
775 rxq->mbufs[i] = NULL;
776 rxq->grant_ref[i] = GRANT_REF_INVALID;
777 }
778
779 /* Start resources allocation */
780
781 if (gnttab_alloc_grant_references(NET_RX_RING_SIZE,
782 &rxq->gref_head) != 0) {
783 device_printf(dev, "allocating rx gref");
784 error = ENOMEM;
785 goto fail;
786 }
787
788 rxs = (netif_rx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
789 M_WAITOK|M_ZERO);
790 SHARED_RING_INIT(rxs);
791 FRONT_RING_INIT(&rxq->ring, rxs, PAGE_SIZE);
792
793 error = xenbus_grant_ring(dev, virt_to_mfn(rxs),
794 &rxq->ring_ref);
795 if (error != 0) {
796 device_printf(dev, "granting rx ring page");
797 goto fail_grant_ring;
798 }
799
800 callout_init(&rxq->rx_refill, 1);
801 }
802
803 return (0);
804
805 fail_grant_ring:
806 gnttab_free_grant_references(rxq->gref_head);
807 free(rxq->ring.sring, M_DEVBUF);
808 fail:
809 for (; q >= 0; q--) {
810 disconnect_rxq(&info->rxq[q]);
811 destroy_rxq(&info->rxq[q]);
812 }
813
814 free(info->rxq, M_DEVBUF);
815 return (error);
816 }
817
818 static void
819 disconnect_txq(struct netfront_txq *txq)
820 {
821
822 xn_release_tx_bufs(txq);
823 gnttab_free_grant_references(txq->gref_head);
824 gnttab_end_foreign_access(txq->ring_ref, NULL);
825 xen_intr_unbind(&txq->xen_intr_handle);
826 }
827
828 static void
829 destroy_txq(struct netfront_txq *txq)
830 {
831 unsigned int i;
832
833 free(txq->ring.sring, M_DEVBUF);
834 buf_ring_free(txq->br, M_DEVBUF);
835 taskqueue_drain_all(txq->tq);
836 taskqueue_free(txq->tq);
837
838 for (i = 0; i <= NET_TX_RING_SIZE; i++) {
839 bus_dmamap_destroy(txq->info->dma_tag,
840 txq->xennet_tag[i].dma_map);
841 txq->xennet_tag[i].dma_map = NULL;
842 }
843 }
844
845 static void
846 destroy_txqs(struct netfront_info *np)
847 {
848 int i;
849
850 for (i = 0; i < np->num_queues; i++)
851 destroy_txq(&np->txq[i]);
852
853 free(np->txq, M_DEVBUF);
854 np->txq = NULL;
855 }
856
857 static int
858 setup_txqs(device_t dev, struct netfront_info *info,
859 unsigned long num_queues)
860 {
861 int q, i;
862 int error;
863 netif_tx_sring_t *txs;
864 struct netfront_txq *txq;
865
866 info->txq = malloc(sizeof(struct netfront_txq) * num_queues,
867 M_DEVBUF, M_WAITOK|M_ZERO);
868
869 for (q = 0; q < num_queues; q++) {
870 txq = &info->txq[q];
871
872 txq->id = q;
873 txq->info = info;
874
875 txq->ring_ref = GRANT_REF_INVALID;
876 txq->ring.sring = NULL;
877
878 snprintf(txq->name, XN_QUEUE_NAME_LEN, "xntx_%u", q);
879
880 mtx_init(&txq->lock, txq->name, "netfront transmit lock",
881 MTX_DEF);
882 SLIST_INIT(&txq->tags);
883
884 for (i = 0; i <= NET_TX_RING_SIZE; i++) {
885 txq->mbufs[i] = (void *) ((u_long) i+1);
886 txq->grant_ref[i] = GRANT_REF_INVALID;
887 txq->xennet_tag[i].txq = txq;
888 txq->xennet_tag[i].dma_tag = info->dma_tag;
889 error = bus_dmamap_create(info->dma_tag, 0,
890 &txq->xennet_tag[i].dma_map);
891 if (error != 0) {
892 device_printf(dev,
893 "failed to allocate dma map\n");
894 goto fail;
895 }
896 m_tag_setup(&txq->xennet_tag[i].tag,
897 MTAG_COOKIE, MTAG_XENNET,
898 sizeof(txq->xennet_tag[i]) -
899 sizeof(txq->xennet_tag[i].tag));
900 txq->xennet_tag[i].tag.m_tag_free = &tag_free;
901 SLIST_INSERT_HEAD(&txq->tags, &txq->xennet_tag[i],
902 next);
903 }
904 txq->mbufs[NET_TX_RING_SIZE] = (void *)0;
905
906 /* Start resources allocation. */
907
908 if (gnttab_alloc_grant_references(NET_TX_RING_SIZE,
909 &txq->gref_head) != 0) {
910 device_printf(dev, "failed to allocate tx grant refs\n");
911 error = ENOMEM;
912 goto fail;
913 }
914
915 txs = (netif_tx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
916 M_WAITOK|M_ZERO);
917 SHARED_RING_INIT(txs);
918 FRONT_RING_INIT(&txq->ring, txs, PAGE_SIZE);
919
920 error = xenbus_grant_ring(dev, virt_to_mfn(txs),
921 &txq->ring_ref);
922 if (error != 0) {
923 device_printf(dev, "failed to grant tx ring\n");
924 goto fail_grant_ring;
925 }
926
927 txq->br = buf_ring_alloc(NET_TX_RING_SIZE, M_DEVBUF,
928 M_WAITOK, &txq->lock);
929 TASK_INIT(&txq->defrtask, 0, xn_txq_tq_deferred, txq);
930
931 txq->tq = taskqueue_create(txq->name, M_WAITOK,
932 taskqueue_thread_enqueue, &txq->tq);
933
934 error = taskqueue_start_threads(&txq->tq, 1, PI_NET,
935 "%s txq %d", device_get_nameunit(dev), txq->id);
936 if (error != 0) {
937 device_printf(dev, "failed to start tx taskq %d\n",
938 txq->id);
939 goto fail_start_thread;
940 }
941
942 error = xen_intr_alloc_and_bind_local_port(dev,
943 xenbus_get_otherend_id(dev), /* filter */ NULL, xn_intr,
944 &info->txq[q], INTR_TYPE_NET | INTR_MPSAFE | INTR_ENTROPY,
945 &txq->xen_intr_handle);
946
947 if (error != 0) {
948 device_printf(dev, "xen_intr_alloc_and_bind_local_port failed\n");
949 goto fail_bind_port;
950 }
951 }
952
953 return (0);
954
955 fail_bind_port:
956 taskqueue_drain_all(txq->tq);
957 fail_start_thread:
958 buf_ring_free(txq->br, M_DEVBUF);
959 taskqueue_free(txq->tq);
960 gnttab_end_foreign_access(txq->ring_ref, NULL);
961 fail_grant_ring:
962 gnttab_free_grant_references(txq->gref_head);
963 free(txq->ring.sring, M_DEVBUF);
964 fail:
965 for (; q >= 0; q--) {
966 disconnect_txq(&info->txq[q]);
967 destroy_txq(&info->txq[q]);
968 }
969
970 free(info->txq, M_DEVBUF);
971 return (error);
972 }
973
974 static int
975 setup_device(device_t dev, struct netfront_info *info,
976 unsigned long num_queues)
977 {
978 int error;
979 int q;
980
981 if (info->txq)
982 destroy_txqs(info);
983
984 if (info->rxq)
985 destroy_rxqs(info);
986
987 info->num_queues = 0;
988
989 error = setup_rxqs(dev, info, num_queues);
990 if (error != 0)
991 goto out;
992 error = setup_txqs(dev, info, num_queues);
993 if (error != 0)
994 goto out;
995
996 info->num_queues = num_queues;
997
998 /* No split event channel at the moment. */
999 for (q = 0; q < num_queues; q++)
1000 info->rxq[q].xen_intr_handle = info->txq[q].xen_intr_handle;
1001
1002 return (0);
1003
1004 out:
1005 KASSERT(error != 0, ("Error path taken without providing an error code"));
1006 return (error);
1007 }
1008
1009 #ifdef INET
1010 /**
1011 * If this interface has an ipv4 address, send an arp for it. This
1012 * helps to get the network going again after migrating hosts.
1013 */
1014 static void
1015 netfront_send_fake_arp(device_t dev, struct netfront_info *info)
1016 {
1017 struct ifnet *ifp;
1018 struct ifaddr *ifa;
1019
1020 ifp = info->xn_ifp;
1021 CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
1022 if (ifa->ifa_addr->sa_family == AF_INET) {
1023 arp_ifinit(ifp, ifa);
1024 }
1025 }
1026 }
1027 #endif
1028
1029 /**
1030 * Callback received when the backend's state changes.
1031 */
1032 static void
1033 netfront_backend_changed(device_t dev, XenbusState newstate)
1034 {
1035 struct netfront_info *sc = device_get_softc(dev);
1036
1037 DPRINTK("newstate=%d\n", newstate);
1038
1039 CURVNET_SET(sc->xn_ifp->if_vnet);
1040
1041 switch (newstate) {
1042 case XenbusStateInitialising:
1043 case XenbusStateInitialised:
1044 case XenbusStateUnknown:
1045 case XenbusStateReconfigured:
1046 case XenbusStateReconfiguring:
1047 break;
1048 case XenbusStateInitWait:
1049 if (xenbus_get_state(dev) != XenbusStateInitialising)
1050 break;
1051 if (xn_connect(sc) != 0)
1052 break;
1053 /* Switch to connected state before kicking the rings. */
1054 xenbus_set_state(sc->xbdev, XenbusStateConnected);
1055 xn_kick_rings(sc);
1056 break;
1057 case XenbusStateClosing:
1058 xenbus_set_state(dev, XenbusStateClosed);
1059 break;
1060 case XenbusStateClosed:
1061 if (sc->xn_reset) {
1062 netif_disconnect_backend(sc);
1063 xenbus_set_state(dev, XenbusStateInitialising);
1064 sc->xn_reset = false;
1065 }
1066 break;
1067 case XenbusStateConnected:
1068 #ifdef INET
1069 netfront_send_fake_arp(dev, sc);
1070 #endif
1071 break;
1072 }
1073
1074 CURVNET_RESTORE();
1075 }
1076
1077 /**
1078 * \brief Verify that there is sufficient space in the Tx ring
1079 * buffer for a maximally sized request to be enqueued.
1080 *
1081 * A transmit request requires a transmit descriptor for each packet
1082 * fragment, plus up to 2 entries for "options" (e.g. TSO).
1083 */
1084 static inline int
1085 xn_tx_slot_available(struct netfront_txq *txq)
1086 {
1087
1088 return (RING_FREE_REQUESTS(&txq->ring) > (MAX_TX_REQ_FRAGS + 2));
1089 }
1090
1091 static void
1092 xn_release_tx_bufs(struct netfront_txq *txq)
1093 {
1094 int i;
1095
1096 for (i = 1; i <= NET_TX_RING_SIZE; i++) {
1097 struct mbuf *m;
1098
1099 m = txq->mbufs[i];
1100
1101 /*
1102 * We assume that no kernel addresses are
1103 * less than NET_TX_RING_SIZE. Any entry
1104 * in the table that is below this number
1105 * must be an index from free-list tracking.
1106 */
1107 if (((uintptr_t)m) <= NET_TX_RING_SIZE)
1108 continue;
1109 gnttab_end_foreign_access_ref(txq->grant_ref[i]);
1110 gnttab_release_grant_reference(&txq->gref_head,
1111 txq->grant_ref[i]);
1112 txq->grant_ref[i] = GRANT_REF_INVALID;
1113 add_id_to_freelist(txq->mbufs, i);
1114 txq->mbufs_cnt--;
1115 if (txq->mbufs_cnt < 0) {
1116 panic("%s: tx_chain_cnt must be >= 0", __func__);
1117 }
1118 mbuf_release(m);
1119 }
1120 }
1121
1122 static struct mbuf *
1123 xn_alloc_one_rx_buffer(struct netfront_rxq *rxq)
1124 {
1125 struct mbuf *m;
1126
1127 m = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, MJUMPAGESIZE);
1128 if (m == NULL)
1129 return NULL;
1130 m->m_len = m->m_pkthdr.len = MJUMPAGESIZE;
1131
1132 return (m);
1133 }
1134
1135 static void
1136 xn_alloc_rx_buffers(struct netfront_rxq *rxq)
1137 {
1138 RING_IDX req_prod;
1139 int notify;
1140
1141 XN_RX_LOCK_ASSERT(rxq);
1142
1143 if (__predict_false(rxq->info->carrier == 0))
1144 return;
1145
1146 for (req_prod = rxq->ring.req_prod_pvt;
1147 req_prod - rxq->ring.rsp_cons < NET_RX_RING_SIZE;
1148 req_prod++) {
1149 struct mbuf *m;
1150 unsigned short id;
1151 grant_ref_t ref;
1152 struct netif_rx_request *req;
1153 unsigned long pfn;
1154
1155 m = xn_alloc_one_rx_buffer(rxq);
1156 if (m == NULL)
1157 break;
1158
1159 id = xn_rxidx(req_prod);
1160
1161 KASSERT(rxq->mbufs[id] == NULL, ("non-NULL xn_rx_chain"));
1162 rxq->mbufs[id] = m;
1163
1164 ref = gnttab_claim_grant_reference(&rxq->gref_head);
1165 KASSERT(ref != GNTTAB_LIST_END,
1166 ("reserved grant references exhuasted"));
1167 rxq->grant_ref[id] = ref;
1168
1169 pfn = atop(vtophys(mtod(m, vm_offset_t)));
1170 req = RING_GET_REQUEST(&rxq->ring, req_prod);
1171
1172 gnttab_grant_foreign_access_ref(ref,
1173 xenbus_get_otherend_id(rxq->info->xbdev), pfn, 0);
1174 req->id = id;
1175 req->gref = ref;
1176 }
1177
1178 rxq->ring.req_prod_pvt = req_prod;
1179
1180 /* Not enough requests? Try again later. */
1181 if (req_prod - rxq->ring.rsp_cons < NET_RX_SLOTS_MIN) {
1182 callout_reset_curcpu(&rxq->rx_refill, hz/10,
1183 xn_alloc_rx_buffers_callout, rxq);
1184 return;
1185 }
1186
1187 wmb(); /* barrier so backend seens requests */
1188
1189 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&rxq->ring, notify);
1190 if (notify)
1191 xen_intr_signal(rxq->xen_intr_handle);
1192 }
1193
1194 static void xn_alloc_rx_buffers_callout(void *arg)
1195 {
1196 struct netfront_rxq *rxq;
1197
1198 rxq = (struct netfront_rxq *)arg;
1199 XN_RX_LOCK(rxq);
1200 xn_alloc_rx_buffers(rxq);
1201 XN_RX_UNLOCK(rxq);
1202 }
1203
1204 static void
1205 xn_release_rx_bufs(struct netfront_rxq *rxq)
1206 {
1207 int i, ref;
1208 struct mbuf *m;
1209
1210 for (i = 0; i < NET_RX_RING_SIZE; i++) {
1211 m = rxq->mbufs[i];
1212
1213 if (m == NULL)
1214 continue;
1215
1216 ref = rxq->grant_ref[i];
1217 if (ref == GRANT_REF_INVALID)
1218 continue;
1219
1220 gnttab_end_foreign_access_ref(ref);
1221 gnttab_release_grant_reference(&rxq->gref_head, ref);
1222 rxq->mbufs[i] = NULL;
1223 rxq->grant_ref[i] = GRANT_REF_INVALID;
1224 m_freem(m);
1225 }
1226 }
1227
1228 static void
1229 xn_rxeof(struct netfront_rxq *rxq)
1230 {
1231 struct ifnet *ifp;
1232 struct netfront_info *np = rxq->info;
1233 #if (defined(INET) || defined(INET6))
1234 struct lro_ctrl *lro = &rxq->lro;
1235 #endif
1236 struct netfront_rx_info rinfo;
1237 struct netif_rx_response *rx = &rinfo.rx;
1238 struct netif_extra_info *extras = rinfo.extras;
1239 RING_IDX i, rp;
1240 struct mbuf *m;
1241 struct mbufq mbufq_rxq, mbufq_errq;
1242 int err, work_to_do;
1243
1244 XN_RX_LOCK_ASSERT(rxq);
1245
1246 if (!netfront_carrier_ok(np))
1247 return;
1248
1249 /* XXX: there should be some sane limit. */
1250 mbufq_init(&mbufq_errq, INT_MAX);
1251 mbufq_init(&mbufq_rxq, INT_MAX);
1252
1253 ifp = np->xn_ifp;
1254
1255 do {
1256 rp = rxq->ring.sring->rsp_prod;
1257 rmb(); /* Ensure we see queued responses up to 'rp'. */
1258
1259 i = rxq->ring.rsp_cons;
1260 while ((i != rp)) {
1261 memcpy(rx, RING_GET_RESPONSE(&rxq->ring, i), sizeof(*rx));
1262 memset(extras, 0, sizeof(rinfo.extras));
1263
1264 m = NULL;
1265 err = xn_get_responses(rxq, &rinfo, rp, &i, &m);
1266
1267 if (__predict_false(err)) {
1268 if (m)
1269 (void )mbufq_enqueue(&mbufq_errq, m);
1270 if_inc_counter(ifp, IFCOUNTER_IQDROPS, 1);
1271 continue;
1272 }
1273
1274 m->m_pkthdr.rcvif = ifp;
1275 if (rx->flags & NETRXF_data_validated) {
1276 /*
1277 * According to mbuf(9) the correct way to tell
1278 * the stack that the checksum of an inbound
1279 * packet is correct, without it actually being
1280 * present (because the underlying interface
1281 * doesn't provide it), is to set the
1282 * CSUM_DATA_VALID and CSUM_PSEUDO_HDR flags,
1283 * and the csum_data field to 0xffff.
1284 */
1285 m->m_pkthdr.csum_flags |= (CSUM_DATA_VALID
1286 | CSUM_PSEUDO_HDR);
1287 m->m_pkthdr.csum_data = 0xffff;
1288 }
1289 if ((rx->flags & NETRXF_extra_info) != 0 &&
1290 (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type ==
1291 XEN_NETIF_EXTRA_TYPE_GSO)) {
1292 m->m_pkthdr.tso_segsz =
1293 extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].u.gso.size;
1294 m->m_pkthdr.csum_flags |= CSUM_TSO;
1295 }
1296
1297 (void )mbufq_enqueue(&mbufq_rxq, m);
1298 }
1299
1300 rxq->ring.rsp_cons = i;
1301
1302 xn_alloc_rx_buffers(rxq);
1303
1304 RING_FINAL_CHECK_FOR_RESPONSES(&rxq->ring, work_to_do);
1305 } while (work_to_do);
1306
1307 mbufq_drain(&mbufq_errq);
1308 /*
1309 * Process all the mbufs after the remapping is complete.
1310 * Break the mbuf chain first though.
1311 */
1312 while ((m = mbufq_dequeue(&mbufq_rxq)) != NULL) {
1313 if_inc_counter(ifp, IFCOUNTER_IPACKETS, 1);
1314 #if (defined(INET) || defined(INET6))
1315 /* Use LRO if possible */
1316 if ((ifp->if_capenable & IFCAP_LRO) == 0 ||
1317 lro->lro_cnt == 0 || tcp_lro_rx(lro, m, 0)) {
1318 /*
1319 * If LRO fails, pass up to the stack
1320 * directly.
1321 */
1322 (*ifp->if_input)(ifp, m);
1323 }
1324 #else
1325 (*ifp->if_input)(ifp, m);
1326 #endif
1327 }
1328
1329 #if (defined(INET) || defined(INET6))
1330 /*
1331 * Flush any outstanding LRO work
1332 */
1333 tcp_lro_flush_all(lro);
1334 #endif
1335 }
1336
1337 static void
1338 xn_txeof(struct netfront_txq *txq)
1339 {
1340 RING_IDX i, prod;
1341 unsigned short id;
1342 struct ifnet *ifp;
1343 netif_tx_response_t *txr;
1344 struct mbuf *m;
1345 struct netfront_info *np = txq->info;
1346
1347 XN_TX_LOCK_ASSERT(txq);
1348
1349 if (!netfront_carrier_ok(np))
1350 return;
1351
1352 ifp = np->xn_ifp;
1353
1354 do {
1355 prod = txq->ring.sring->rsp_prod;
1356 rmb(); /* Ensure we see responses up to 'rp'. */
1357
1358 for (i = txq->ring.rsp_cons; i != prod; i++) {
1359 txr = RING_GET_RESPONSE(&txq->ring, i);
1360 if (txr->status == NETIF_RSP_NULL)
1361 continue;
1362
1363 if (txr->status != NETIF_RSP_OKAY) {
1364 printf("%s: WARNING: response is %d!\n",
1365 __func__, txr->status);
1366 }
1367 id = txr->id;
1368 m = txq->mbufs[id];
1369 KASSERT(m != NULL, ("mbuf not found in chain"));
1370 KASSERT((uintptr_t)m > NET_TX_RING_SIZE,
1371 ("mbuf already on the free list, but we're "
1372 "trying to free it again!"));
1373 M_ASSERTVALID(m);
1374
1375 if (__predict_false(gnttab_query_foreign_access(
1376 txq->grant_ref[id]) != 0)) {
1377 panic("%s: grant id %u still in use by the "
1378 "backend", __func__, id);
1379 }
1380 gnttab_end_foreign_access_ref(txq->grant_ref[id]);
1381 gnttab_release_grant_reference(
1382 &txq->gref_head, txq->grant_ref[id]);
1383 txq->grant_ref[id] = GRANT_REF_INVALID;
1384
1385 txq->mbufs[id] = NULL;
1386 add_id_to_freelist(txq->mbufs, id);
1387 txq->mbufs_cnt--;
1388 mbuf_release(m);
1389 /* Only mark the txq active if we've freed up at least one slot to try */
1390 ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1391 }
1392 txq->ring.rsp_cons = prod;
1393
1394 /*
1395 * Set a new event, then check for race with update of
1396 * tx_cons. Note that it is essential to schedule a
1397 * callback, no matter how few buffers are pending. Even if
1398 * there is space in the transmit ring, higher layers may
1399 * be blocked because too much data is outstanding: in such
1400 * cases notification from Xen is likely to be the only kick
1401 * that we'll get.
1402 */
1403 txq->ring.sring->rsp_event =
1404 prod + ((txq->ring.sring->req_prod - prod) >> 1) + 1;
1405
1406 mb();
1407 } while (prod != txq->ring.sring->rsp_prod);
1408
1409 if (txq->full &&
1410 ((txq->ring.sring->req_prod - prod) < NET_TX_RING_SIZE)) {
1411 txq->full = false;
1412 xn_txq_start(txq);
1413 }
1414 }
1415
1416 static void
1417 xn_intr(void *xsc)
1418 {
1419 struct netfront_txq *txq = xsc;
1420 struct netfront_info *np = txq->info;
1421 struct netfront_rxq *rxq = &np->rxq[txq->id];
1422
1423 /* kick both tx and rx */
1424 xn_rxq_intr(rxq);
1425 xn_txq_intr(txq);
1426 }
1427
1428 static void
1429 xn_move_rx_slot(struct netfront_rxq *rxq, struct mbuf *m,
1430 grant_ref_t ref)
1431 {
1432 int new = xn_rxidx(rxq->ring.req_prod_pvt);
1433
1434 KASSERT(rxq->mbufs[new] == NULL, ("mbufs != NULL"));
1435 rxq->mbufs[new] = m;
1436 rxq->grant_ref[new] = ref;
1437 RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->id = new;
1438 RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->gref = ref;
1439 rxq->ring.req_prod_pvt++;
1440 }
1441
1442 static int
1443 xn_get_extras(struct netfront_rxq *rxq,
1444 struct netif_extra_info *extras, RING_IDX rp, RING_IDX *cons)
1445 {
1446 struct netif_extra_info *extra;
1447
1448 int err = 0;
1449
1450 do {
1451 struct mbuf *m;
1452 grant_ref_t ref;
1453
1454 if (__predict_false(*cons + 1 == rp)) {
1455 err = EINVAL;
1456 break;
1457 }
1458
1459 extra = (struct netif_extra_info *)
1460 RING_GET_RESPONSE(&rxq->ring, ++(*cons));
1461
1462 if (__predict_false(!extra->type ||
1463 extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1464 err = EINVAL;
1465 } else {
1466 memcpy(&extras[extra->type - 1], extra, sizeof(*extra));
1467 }
1468
1469 m = xn_get_rx_mbuf(rxq, *cons);
1470 ref = xn_get_rx_ref(rxq, *cons);
1471 xn_move_rx_slot(rxq, m, ref);
1472 } while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
1473
1474 return err;
1475 }
1476
1477 static int
1478 xn_get_responses(struct netfront_rxq *rxq,
1479 struct netfront_rx_info *rinfo, RING_IDX rp, RING_IDX *cons,
1480 struct mbuf **list)
1481 {
1482 struct netif_rx_response *rx = &rinfo->rx;
1483 struct netif_extra_info *extras = rinfo->extras;
1484 struct mbuf *m, *m0, *m_prev;
1485 grant_ref_t ref = xn_get_rx_ref(rxq, *cons);
1486 int frags = 1;
1487 int err = 0;
1488 u_long ret __diagused;
1489
1490 m0 = m = m_prev = xn_get_rx_mbuf(rxq, *cons);
1491
1492 if (rx->flags & NETRXF_extra_info) {
1493 err = xn_get_extras(rxq, extras, rp, cons);
1494 }
1495
1496 if (m0 != NULL) {
1497 m0->m_pkthdr.len = 0;
1498 m0->m_next = NULL;
1499 }
1500
1501 for (;;) {
1502 #if 0
1503 DPRINTK("rx->status=%hd rx->offset=%hu frags=%u\n",
1504 rx->status, rx->offset, frags);
1505 #endif
1506 if (__predict_false(rx->status < 0 ||
1507 rx->offset + rx->status > PAGE_SIZE)) {
1508 xn_move_rx_slot(rxq, m, ref);
1509 if (m0 == m)
1510 m0 = NULL;
1511 m = NULL;
1512 err = EINVAL;
1513 goto next_skip_queue;
1514 }
1515
1516 /*
1517 * This definitely indicates a bug, either in this driver or in
1518 * the backend driver. In future this should flag the bad
1519 * situation to the system controller to reboot the backed.
1520 */
1521 if (ref == GRANT_REF_INVALID) {
1522 printf("%s: Bad rx response id %d.\n", __func__, rx->id);
1523 err = EINVAL;
1524 goto next;
1525 }
1526
1527 ret = gnttab_end_foreign_access_ref(ref);
1528 KASSERT(ret, ("Unable to end access to grant references"));
1529
1530 gnttab_release_grant_reference(&rxq->gref_head, ref);
1531
1532 next:
1533 if (m == NULL)
1534 break;
1535
1536 m->m_len = rx->status;
1537 m->m_data += rx->offset;
1538 m0->m_pkthdr.len += rx->status;
1539
1540 next_skip_queue:
1541 if (!(rx->flags & NETRXF_more_data))
1542 break;
1543
1544 if (*cons + frags == rp) {
1545 if (net_ratelimit())
1546 WPRINTK("Need more frags\n");
1547 err = ENOENT;
1548 printf("%s: cons %u frags %u rp %u, not enough frags\n",
1549 __func__, *cons, frags, rp);
1550 break;
1551 }
1552 /*
1553 * Note that m can be NULL, if rx->status < 0 or if
1554 * rx->offset + rx->status > PAGE_SIZE above.
1555 */
1556 m_prev = m;
1557
1558 rx = RING_GET_RESPONSE(&rxq->ring, *cons + frags);
1559 m = xn_get_rx_mbuf(rxq, *cons + frags);
1560
1561 /*
1562 * m_prev == NULL can happen if rx->status < 0 or if
1563 * rx->offset + * rx->status > PAGE_SIZE above.
1564 */
1565 if (m_prev != NULL)
1566 m_prev->m_next = m;
1567
1568 /*
1569 * m0 can be NULL if rx->status < 0 or if * rx->offset +
1570 * rx->status > PAGE_SIZE above.
1571 */
1572 if (m0 == NULL)
1573 m0 = m;
1574 m->m_next = NULL;
1575 ref = xn_get_rx_ref(rxq, *cons + frags);
1576 frags++;
1577 }
1578 *list = m0;
1579 *cons += frags;
1580
1581 return (err);
1582 }
1583
1584 /**
1585 * Given an mbuf chain, make sure we have enough room and then push
1586 * it onto the transmit ring.
1587 */
1588 static int
1589 xn_assemble_tx_request(struct netfront_txq *txq, struct mbuf *m_head)
1590 {
1591 struct netfront_info *np = txq->info;
1592 struct ifnet *ifp = np->xn_ifp;
1593 int otherend_id, error, nfrags;
1594 bus_dma_segment_t *segs = txq->segs;
1595 struct mbuf_xennet *tag;
1596 bus_dmamap_t map;
1597 unsigned int i;
1598
1599 KASSERT(!SLIST_EMPTY(&txq->tags), ("no tags available"));
1600 tag = SLIST_FIRST(&txq->tags);
1601 SLIST_REMOVE_HEAD(&txq->tags, next);
1602 KASSERT(tag->count == 0, ("tag already in-use"));
1603 map = tag->dma_map;
1604 error = bus_dmamap_load_mbuf_sg(np->dma_tag, map, m_head, segs,
1605 &nfrags, 0);
1606 if (error == EFBIG || nfrags > np->maxfrags) {
1607 struct mbuf *m;
1608
1609 bus_dmamap_unload(np->dma_tag, map);
1610 m = m_defrag(m_head, M_NOWAIT);
1611 if (!m) {
1612 /*
1613 * Defrag failed, so free the mbuf and
1614 * therefore drop the packet.
1615 */
1616 SLIST_INSERT_HEAD(&txq->tags, tag, next);
1617 m_freem(m_head);
1618 return (EMSGSIZE);
1619 }
1620 m_head = m;
1621 error = bus_dmamap_load_mbuf_sg(np->dma_tag, map, m_head, segs,
1622 &nfrags, 0);
1623 if (error != 0 || nfrags > np->maxfrags) {
1624 bus_dmamap_unload(np->dma_tag, map);
1625 SLIST_INSERT_HEAD(&txq->tags, tag, next);
1626 m_freem(m_head);
1627 return (error ?: EFBIG);
1628 }
1629 } else if (error != 0) {
1630 SLIST_INSERT_HEAD(&txq->tags, tag, next);
1631 m_freem(m_head);
1632 return (error);
1633 }
1634
1635 /**
1636 * The FreeBSD TCP stack, with TSO enabled, can produce a chain
1637 * of mbufs longer than Linux can handle. Make sure we don't
1638 * pass a too-long chain over to the other side by dropping the
1639 * packet. It doesn't look like there is currently a way to
1640 * tell the TCP stack to generate a shorter chain of packets.
1641 */
1642 if (nfrags > MAX_TX_REQ_FRAGS) {
1643 #ifdef DEBUG
1644 printf("%s: nfrags %d > MAX_TX_REQ_FRAGS %d, netback "
1645 "won't be able to handle it, dropping\n",
1646 __func__, nfrags, MAX_TX_REQ_FRAGS);
1647 #endif
1648 SLIST_INSERT_HEAD(&txq->tags, tag, next);
1649 bus_dmamap_unload(np->dma_tag, map);
1650 m_freem(m_head);
1651 return (EMSGSIZE);
1652 }
1653
1654 /*
1655 * This check should be redundant. We've already verified that we
1656 * have enough slots in the ring to handle a packet of maximum
1657 * size, and that our packet is less than the maximum size. Keep
1658 * it in here as an assert for now just to make certain that
1659 * chain_cnt is accurate.
1660 */
1661 KASSERT((txq->mbufs_cnt + nfrags) <= NET_TX_RING_SIZE,
1662 ("%s: chain_cnt (%d) + nfrags (%d) > NET_TX_RING_SIZE "
1663 "(%d)!", __func__, (int) txq->mbufs_cnt,
1664 (int) nfrags, (int) NET_TX_RING_SIZE));
1665
1666 /*
1667 * Start packing the mbufs in this chain into
1668 * the fragment pointers. Stop when we run out
1669 * of fragments or hit the end of the mbuf chain.
1670 */
1671 otherend_id = xenbus_get_otherend_id(np->xbdev);
1672 m_tag_prepend(m_head, &tag->tag);
1673 for (i = 0; i < nfrags; i++) {
1674 netif_tx_request_t *tx;
1675 uintptr_t id;
1676 grant_ref_t ref;
1677 u_long mfn; /* XXX Wrong type? */
1678
1679 tx = RING_GET_REQUEST(&txq->ring, txq->ring.req_prod_pvt);
1680 id = get_id_from_freelist(txq->mbufs);
1681 if (id == 0)
1682 panic("%s: was allocated the freelist head!\n",
1683 __func__);
1684 txq->mbufs_cnt++;
1685 if (txq->mbufs_cnt > NET_TX_RING_SIZE)
1686 panic("%s: tx_chain_cnt must be <= NET_TX_RING_SIZE\n",
1687 __func__);
1688 mbuf_grab(m_head);
1689 txq->mbufs[id] = m_head;
1690 tx->id = id;
1691 ref = gnttab_claim_grant_reference(&txq->gref_head);
1692 KASSERT((short)ref >= 0, ("Negative ref"));
1693 mfn = atop(segs[i].ds_addr);
1694 gnttab_grant_foreign_access_ref(ref, otherend_id,
1695 mfn, GNTMAP_readonly);
1696 tx->gref = txq->grant_ref[id] = ref;
1697 tx->offset = segs[i].ds_addr & PAGE_MASK;
1698 KASSERT(tx->offset + segs[i].ds_len <= PAGE_SIZE,
1699 ("mbuf segment crosses a page boundary"));
1700 tx->flags = 0;
1701 if (i == 0) {
1702 /*
1703 * The first fragment has the entire packet
1704 * size, subsequent fragments have just the
1705 * fragment size. The backend works out the
1706 * true size of the first fragment by
1707 * subtracting the sizes of the other
1708 * fragments.
1709 */
1710 tx->size = m_head->m_pkthdr.len;
1711
1712 /*
1713 * The first fragment contains the checksum flags
1714 * and is optionally followed by extra data for
1715 * TSO etc.
1716 */
1717 /**
1718 * CSUM_TSO requires checksum offloading.
1719 * Some versions of FreeBSD fail to
1720 * set CSUM_TCP in the CSUM_TSO case,
1721 * so we have to test for CSUM_TSO
1722 * explicitly.
1723 */
1724 if (m_head->m_pkthdr.csum_flags
1725 & (CSUM_DELAY_DATA | CSUM_TSO)) {
1726 tx->flags |= (NETTXF_csum_blank
1727 | NETTXF_data_validated);
1728 }
1729 if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
1730 struct netif_extra_info *gso =
1731 (struct netif_extra_info *)
1732 RING_GET_REQUEST(&txq->ring,
1733 ++txq->ring.req_prod_pvt);
1734
1735 tx->flags |= NETTXF_extra_info;
1736
1737 gso->u.gso.size = m_head->m_pkthdr.tso_segsz;
1738 gso->u.gso.type =
1739 XEN_NETIF_GSO_TYPE_TCPV4;
1740 gso->u.gso.pad = 0;
1741 gso->u.gso.features = 0;
1742
1743 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
1744 gso->flags = 0;
1745 }
1746 } else {
1747 tx->size = segs[i].ds_len;
1748 }
1749 if (i != nfrags - 1)
1750 tx->flags |= NETTXF_more_data;
1751
1752 txq->ring.req_prod_pvt++;
1753 }
1754 bus_dmamap_sync(np->dma_tag, map, BUS_DMASYNC_PREWRITE);
1755 BPF_MTAP(ifp, m_head);
1756
1757 if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
1758 if_inc_counter(ifp, IFCOUNTER_OBYTES, m_head->m_pkthdr.len);
1759 if (m_head->m_flags & M_MCAST)
1760 if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
1761
1762 xn_txeof(txq);
1763
1764 return (0);
1765 }
1766
1767 /* equivalent of network_open() in Linux */
1768 static void
1769 xn_ifinit_locked(struct netfront_info *np)
1770 {
1771 struct ifnet *ifp;
1772 int i;
1773 struct netfront_rxq *rxq;
1774
1775 XN_LOCK_ASSERT(np);
1776
1777 ifp = np->xn_ifp;
1778
1779 if (ifp->if_drv_flags & IFF_DRV_RUNNING || !netfront_carrier_ok(np))
1780 return;
1781
1782 xn_stop(np);
1783
1784 for (i = 0; i < np->num_queues; i++) {
1785 rxq = &np->rxq[i];
1786 XN_RX_LOCK(rxq);
1787 xn_alloc_rx_buffers(rxq);
1788 rxq->ring.sring->rsp_event = rxq->ring.rsp_cons + 1;
1789 if (RING_HAS_UNCONSUMED_RESPONSES(&rxq->ring))
1790 xn_rxeof(rxq);
1791 XN_RX_UNLOCK(rxq);
1792 }
1793
1794 ifp->if_drv_flags |= IFF_DRV_RUNNING;
1795 ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1796 if_link_state_change(ifp, LINK_STATE_UP);
1797 }
1798
1799 static void
1800 xn_ifinit(void *xsc)
1801 {
1802 struct netfront_info *sc = xsc;
1803
1804 XN_LOCK(sc);
1805 xn_ifinit_locked(sc);
1806 XN_UNLOCK(sc);
1807 }
1808
1809 static int
1810 xn_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
1811 {
1812 struct netfront_info *sc = ifp->if_softc;
1813 struct ifreq *ifr = (struct ifreq *) data;
1814 device_t dev;
1815 #ifdef INET
1816 struct ifaddr *ifa = (struct ifaddr *)data;
1817 #endif
1818 int mask, error = 0, reinit;
1819
1820 dev = sc->xbdev;
1821
1822 switch(cmd) {
1823 case SIOCSIFADDR:
1824 #ifdef INET
1825 XN_LOCK(sc);
1826 if (ifa->ifa_addr->sa_family == AF_INET) {
1827 ifp->if_flags |= IFF_UP;
1828 if (!(ifp->if_drv_flags & IFF_DRV_RUNNING))
1829 xn_ifinit_locked(sc);
1830 arp_ifinit(ifp, ifa);
1831 XN_UNLOCK(sc);
1832 } else {
1833 XN_UNLOCK(sc);
1834 #endif
1835 error = ether_ioctl(ifp, cmd, data);
1836 #ifdef INET
1837 }
1838 #endif
1839 break;
1840 case SIOCSIFMTU:
1841 if (ifp->if_mtu == ifr->ifr_mtu)
1842 break;
1843
1844 ifp->if_mtu = ifr->ifr_mtu;
1845 ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
1846 xn_ifinit(sc);
1847 break;
1848 case SIOCSIFFLAGS:
1849 XN_LOCK(sc);
1850 if (ifp->if_flags & IFF_UP) {
1851 /*
1852 * If only the state of the PROMISC flag changed,
1853 * then just use the 'set promisc mode' command
1854 * instead of reinitializing the entire NIC. Doing
1855 * a full re-init means reloading the firmware and
1856 * waiting for it to start up, which may take a
1857 * second or two.
1858 */
1859 xn_ifinit_locked(sc);
1860 } else {
1861 if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
1862 xn_stop(sc);
1863 }
1864 }
1865 sc->xn_if_flags = ifp->if_flags;
1866 XN_UNLOCK(sc);
1867 break;
1868 case SIOCSIFCAP:
1869 mask = ifr->ifr_reqcap ^ ifp->if_capenable;
1870 reinit = 0;
1871
1872 if (mask & IFCAP_TXCSUM) {
1873 ifp->if_capenable ^= IFCAP_TXCSUM;
1874 ifp->if_hwassist ^= XN_CSUM_FEATURES;
1875 }
1876 if (mask & IFCAP_TSO4) {
1877 ifp->if_capenable ^= IFCAP_TSO4;
1878 ifp->if_hwassist ^= CSUM_TSO;
1879 }
1880
1881 if (mask & (IFCAP_RXCSUM | IFCAP_LRO)) {
1882 /* These Rx features require us to renegotiate. */
1883 reinit = 1;
1884
1885 if (mask & IFCAP_RXCSUM)
1886 ifp->if_capenable ^= IFCAP_RXCSUM;
1887 if (mask & IFCAP_LRO)
1888 ifp->if_capenable ^= IFCAP_LRO;
1889 }
1890
1891 if (reinit == 0)
1892 break;
1893
1894 /*
1895 * We must reset the interface so the backend picks up the
1896 * new features.
1897 */
1898 device_printf(sc->xbdev,
1899 "performing interface reset due to feature change\n");
1900 XN_LOCK(sc);
1901 netfront_carrier_off(sc);
1902 sc->xn_reset = true;
1903 /*
1904 * NB: the pending packet queue is not flushed, since
1905 * the interface should still support the old options.
1906 */
1907 XN_UNLOCK(sc);
1908 /*
1909 * Delete the xenstore nodes that export features.
1910 *
1911 * NB: There's a xenbus state called
1912 * "XenbusStateReconfiguring", which is what we should set
1913 * here. Sadly none of the backends know how to handle it,
1914 * and simply disconnect from the frontend, so we will just
1915 * switch back to XenbusStateInitialising in order to force
1916 * a reconnection.
1917 */
1918 xs_rm(XST_NIL, xenbus_get_node(dev), "feature-gso-tcpv4");
1919 xs_rm(XST_NIL, xenbus_get_node(dev), "feature-no-csum-offload");
1920 xenbus_set_state(dev, XenbusStateClosing);
1921
1922 /*
1923 * Wait for the frontend to reconnect before returning
1924 * from the ioctl. 30s should be more than enough for any
1925 * sane backend to reconnect.
1926 */
1927 error = tsleep(sc, 0, "xn_rst", 30*hz);
1928 break;
1929 case SIOCADDMULTI:
1930 case SIOCDELMULTI:
1931 break;
1932 case SIOCSIFMEDIA:
1933 case SIOCGIFMEDIA:
1934 error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
1935 break;
1936 default:
1937 error = ether_ioctl(ifp, cmd, data);
1938 }
1939
1940 return (error);
1941 }
1942
1943 static void
1944 xn_stop(struct netfront_info *sc)
1945 {
1946 struct ifnet *ifp;
1947
1948 XN_LOCK_ASSERT(sc);
1949
1950 ifp = sc->xn_ifp;
1951
1952 ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
1953 if_link_state_change(ifp, LINK_STATE_DOWN);
1954 }
1955
1956 static void
1957 xn_rebuild_rx_bufs(struct netfront_rxq *rxq)
1958 {
1959 int requeue_idx, i;
1960 grant_ref_t ref;
1961 netif_rx_request_t *req;
1962
1963 for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
1964 struct mbuf *m;
1965 u_long pfn;
1966
1967 if (rxq->mbufs[i] == NULL)
1968 continue;
1969
1970 m = rxq->mbufs[requeue_idx] = xn_get_rx_mbuf(rxq, i);
1971 ref = rxq->grant_ref[requeue_idx] = xn_get_rx_ref(rxq, i);
1972
1973 req = RING_GET_REQUEST(&rxq->ring, requeue_idx);
1974 pfn = vtophys(mtod(m, vm_offset_t)) >> PAGE_SHIFT;
1975
1976 gnttab_grant_foreign_access_ref(ref,
1977 xenbus_get_otherend_id(rxq->info->xbdev),
1978 pfn, 0);
1979
1980 req->gref = ref;
1981 req->id = requeue_idx;
1982
1983 requeue_idx++;
1984 }
1985
1986 rxq->ring.req_prod_pvt = requeue_idx;
1987 }
1988
1989 /* START of Xenolinux helper functions adapted to FreeBSD */
1990 static int
1991 xn_connect(struct netfront_info *np)
1992 {
1993 int i, error;
1994 u_int feature_rx_copy;
1995 struct netfront_rxq *rxq;
1996 struct netfront_txq *txq;
1997
1998 error = xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
1999 "feature-rx-copy", NULL, "%u", &feature_rx_copy);
2000 if (error != 0)
2001 feature_rx_copy = 0;
2002
2003 /* We only support rx copy. */
2004 if (!feature_rx_copy)
2005 return (EPROTONOSUPPORT);
2006
2007 /* Recovery procedure: */
2008 error = talk_to_backend(np->xbdev, np);
2009 if (error != 0)
2010 return (error);
2011
2012 /* Step 1: Reinitialise variables. */
2013 xn_query_features(np);
2014 xn_configure_features(np);
2015
2016 /* Step 2: Release TX buffer */
2017 for (i = 0; i < np->num_queues; i++) {
2018 txq = &np->txq[i];
2019 xn_release_tx_bufs(txq);
2020 }
2021
2022 /* Step 3: Rebuild the RX buffer freelist and the RX ring itself. */
2023 for (i = 0; i < np->num_queues; i++) {
2024 rxq = &np->rxq[i];
2025 xn_rebuild_rx_bufs(rxq);
2026 }
2027
2028 /* Step 4: All public and private state should now be sane. Get
2029 * ready to start sending and receiving packets and give the driver
2030 * domain a kick because we've probably just requeued some
2031 * packets.
2032 */
2033 netfront_carrier_on(np);
2034 wakeup(np);
2035
2036 return (0);
2037 }
2038
2039 static void
2040 xn_kick_rings(struct netfront_info *np)
2041 {
2042 struct netfront_rxq *rxq;
2043 struct netfront_txq *txq;
2044 int i;
2045
2046 for (i = 0; i < np->num_queues; i++) {
2047 txq = &np->txq[i];
2048 rxq = &np->rxq[i];
2049 xen_intr_signal(txq->xen_intr_handle);
2050 XN_TX_LOCK(txq);
2051 xn_txeof(txq);
2052 XN_TX_UNLOCK(txq);
2053 XN_RX_LOCK(rxq);
2054 xn_alloc_rx_buffers(rxq);
2055 XN_RX_UNLOCK(rxq);
2056 }
2057 }
2058
2059 static void
2060 xn_query_features(struct netfront_info *np)
2061 {
2062 int val;
2063
2064 device_printf(np->xbdev, "backend features:");
2065
2066 if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2067 "feature-sg", NULL, "%d", &val) != 0)
2068 val = 0;
2069
2070 np->maxfrags = 1;
2071 if (val) {
2072 np->maxfrags = MAX_TX_REQ_FRAGS;
2073 printf(" feature-sg");
2074 }
2075
2076 if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2077 "feature-gso-tcpv4", NULL, "%d", &val) != 0)
2078 val = 0;
2079
2080 np->xn_ifp->if_capabilities &= ~(IFCAP_TSO4|IFCAP_LRO);
2081 if (val) {
2082 np->xn_ifp->if_capabilities |= IFCAP_TSO4|IFCAP_LRO;
2083 printf(" feature-gso-tcp4");
2084 }
2085
2086 /*
2087 * HW CSUM offload is assumed to be available unless
2088 * feature-no-csum-offload is set in xenstore.
2089 */
2090 if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2091 "feature-no-csum-offload", NULL, "%d", &val) != 0)
2092 val = 0;
2093
2094 np->xn_ifp->if_capabilities |= IFCAP_HWCSUM;
2095 if (val) {
2096 np->xn_ifp->if_capabilities &= ~(IFCAP_HWCSUM);
2097 printf(" feature-no-csum-offload");
2098 }
2099
2100 printf("\n");
2101 }
2102
2103 static int
2104 xn_configure_features(struct netfront_info *np)
2105 {
2106 int err, cap_enabled;
2107 #if (defined(INET) || defined(INET6))
2108 int i;
2109 #endif
2110 struct ifnet *ifp;
2111
2112 ifp = np->xn_ifp;
2113 err = 0;
2114
2115 if ((ifp->if_capenable & ifp->if_capabilities) == ifp->if_capenable) {
2116 /* Current options are available, no need to do anything. */
2117 return (0);
2118 }
2119
2120 /* Try to preserve as many options as possible. */
2121 cap_enabled = ifp->if_capenable;
2122 ifp->if_capenable = ifp->if_hwassist = 0;
2123
2124 #if (defined(INET) || defined(INET6))
2125 if ((cap_enabled & IFCAP_LRO) != 0)
2126 for (i = 0; i < np->num_queues; i++)
2127 tcp_lro_free(&np->rxq[i].lro);
2128 if (xn_enable_lro &&
2129 (ifp->if_capabilities & cap_enabled & IFCAP_LRO) != 0) {
2130 ifp->if_capenable |= IFCAP_LRO;
2131 for (i = 0; i < np->num_queues; i++) {
2132 err = tcp_lro_init(&np->rxq[i].lro);
2133 if (err != 0) {
2134 device_printf(np->xbdev,
2135 "LRO initialization failed\n");
2136 ifp->if_capenable &= ~IFCAP_LRO;
2137 break;
2138 }
2139 np->rxq[i].lro.ifp = ifp;
2140 }
2141 }
2142 if ((ifp->if_capabilities & cap_enabled & IFCAP_TSO4) != 0) {
2143 ifp->if_capenable |= IFCAP_TSO4;
2144 ifp->if_hwassist |= CSUM_TSO;
2145 }
2146 #endif
2147 if ((ifp->if_capabilities & cap_enabled & IFCAP_TXCSUM) != 0) {
2148 ifp->if_capenable |= IFCAP_TXCSUM;
2149 ifp->if_hwassist |= XN_CSUM_FEATURES;
2150 }
2151 if ((ifp->if_capabilities & cap_enabled & IFCAP_RXCSUM) != 0)
2152 ifp->if_capenable |= IFCAP_RXCSUM;
2153
2154 return (err);
2155 }
2156
2157 static int
2158 xn_txq_mq_start_locked(struct netfront_txq *txq, struct mbuf *m)
2159 {
2160 struct netfront_info *np;
2161 struct ifnet *ifp;
2162 struct buf_ring *br;
2163 int error, notify;
2164
2165 np = txq->info;
2166 br = txq->br;
2167 ifp = np->xn_ifp;
2168 error = 0;
2169
2170 XN_TX_LOCK_ASSERT(txq);
2171
2172 if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
2173 !netfront_carrier_ok(np)) {
2174 if (m != NULL)
2175 error = drbr_enqueue(ifp, br, m);
2176 return (error);
2177 }
2178
2179 if (m != NULL) {
2180 error = drbr_enqueue(ifp, br, m);
2181 if (error != 0)
2182 return (error);
2183 }
2184
2185 while ((m = drbr_peek(ifp, br)) != NULL) {
2186 if (!xn_tx_slot_available(txq)) {
2187 drbr_putback(ifp, br, m);
2188 break;
2189 }
2190
2191 error = xn_assemble_tx_request(txq, m);
2192 /* xn_assemble_tx_request always consumes the mbuf*/
2193 if (error != 0) {
2194 drbr_advance(ifp, br);
2195 break;
2196 }
2197
2198 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&txq->ring, notify);
2199 if (notify)
2200 xen_intr_signal(txq->xen_intr_handle);
2201
2202 drbr_advance(ifp, br);
2203 }
2204
2205 if (RING_FULL(&txq->ring))
2206 txq->full = true;
2207
2208 return (0);
2209 }
2210
2211 static int
2212 xn_txq_mq_start(struct ifnet *ifp, struct mbuf *m)
2213 {
2214 struct netfront_info *np;
2215 struct netfront_txq *txq;
2216 int i, npairs, error;
2217
2218 np = ifp->if_softc;
2219 npairs = np->num_queues;
2220
2221 if (!netfront_carrier_ok(np))
2222 return (ENOBUFS);
2223
2224 KASSERT(npairs != 0, ("called with 0 available queues"));
2225
2226 /* check if flowid is set */
2227 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
2228 i = m->m_pkthdr.flowid % npairs;
2229 else
2230 i = curcpu % npairs;
2231
2232 txq = &np->txq[i];
2233
2234 if (XN_TX_TRYLOCK(txq) != 0) {
2235 error = xn_txq_mq_start_locked(txq, m);
2236 XN_TX_UNLOCK(txq);
2237 } else {
2238 error = drbr_enqueue(ifp, txq->br, m);
2239 taskqueue_enqueue(txq->tq, &txq->defrtask);
2240 }
2241
2242 return (error);
2243 }
2244
2245 static void
2246 xn_qflush(struct ifnet *ifp)
2247 {
2248 struct netfront_info *np;
2249 struct netfront_txq *txq;
2250 struct mbuf *m;
2251 int i;
2252
2253 np = ifp->if_softc;
2254
2255 for (i = 0; i < np->num_queues; i++) {
2256 txq = &np->txq[i];
2257
2258 XN_TX_LOCK(txq);
2259 while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
2260 m_freem(m);
2261 XN_TX_UNLOCK(txq);
2262 }
2263
2264 if_qflush(ifp);
2265 }
2266
2267 /**
2268 * Create a network device.
2269 * @param dev Newbus device representing this virtual NIC.
2270 */
2271 int
2272 create_netdev(device_t dev)
2273 {
2274 struct netfront_info *np;
2275 int err;
2276 struct ifnet *ifp;
2277
2278 np = device_get_softc(dev);
2279
2280 np->xbdev = dev;
2281
2282 mtx_init(&np->sc_lock, "xnsc", "netfront softc lock", MTX_DEF);
2283
2284 ifmedia_init(&np->sc_media, 0, xn_ifmedia_upd, xn_ifmedia_sts);
2285 ifmedia_add(&np->sc_media, IFM_ETHER|IFM_MANUAL, 0, NULL);
2286 ifmedia_set(&np->sc_media, IFM_ETHER|IFM_MANUAL);
2287
2288 err = xen_net_read_mac(dev, np->mac);
2289 if (err != 0)
2290 goto error;
2291
2292 /* Set up ifnet structure */
2293 ifp = np->xn_ifp = if_alloc(IFT_ETHER);
2294 ifp->if_softc = np;
2295 if_initname(ifp, "xn", device_get_unit(dev));
2296 ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
2297 ifp->if_ioctl = xn_ioctl;
2298
2299 ifp->if_transmit = xn_txq_mq_start;
2300 ifp->if_qflush = xn_qflush;
2301
2302 ifp->if_init = xn_ifinit;
2303
2304 ifp->if_hwassist = XN_CSUM_FEATURES;
2305 /* Enable all supported features at device creation. */
2306 ifp->if_capenable = ifp->if_capabilities =
2307 IFCAP_HWCSUM|IFCAP_TSO4|IFCAP_LRO;
2308 ifp->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
2309 ifp->if_hw_tsomaxsegcount = MAX_TX_REQ_FRAGS;
2310 ifp->if_hw_tsomaxsegsize = PAGE_SIZE;
2311
2312 ether_ifattach(ifp, np->mac);
2313 netfront_carrier_off(np);
2314
2315 err = bus_dma_tag_create(
2316 bus_get_dma_tag(dev), /* parent */
2317 1, PAGE_SIZE, /* algnmnt, boundary */
2318 BUS_SPACE_MAXADDR, /* lowaddr */
2319 BUS_SPACE_MAXADDR, /* highaddr */
2320 NULL, NULL, /* filter, filterarg */
2321 PAGE_SIZE * MAX_TX_REQ_FRAGS, /* max request size */
2322 MAX_TX_REQ_FRAGS, /* max segments */
2323 PAGE_SIZE, /* maxsegsize */
2324 BUS_DMA_ALLOCNOW, /* flags */
2325 NULL, NULL, /* lockfunc, lockarg */
2326 &np->dma_tag);
2327
2328 return (err);
2329
2330 error:
2331 KASSERT(err != 0, ("Error path with no error code specified"));
2332 return (err);
2333 }
2334
2335 static int
2336 netfront_detach(device_t dev)
2337 {
2338 struct netfront_info *info = device_get_softc(dev);
2339
2340 DPRINTK("%s\n", xenbus_get_node(dev));
2341
2342 netif_free(info);
2343
2344 return 0;
2345 }
2346
2347 static void
2348 netif_free(struct netfront_info *np)
2349 {
2350
2351 XN_LOCK(np);
2352 xn_stop(np);
2353 XN_UNLOCK(np);
2354 netif_disconnect_backend(np);
2355 ether_ifdetach(np->xn_ifp);
2356 free(np->rxq, M_DEVBUF);
2357 free(np->txq, M_DEVBUF);
2358 if_free(np->xn_ifp);
2359 np->xn_ifp = NULL;
2360 ifmedia_removeall(&np->sc_media);
2361 bus_dma_tag_destroy(np->dma_tag);
2362 }
2363
2364 static void
2365 netif_disconnect_backend(struct netfront_info *np)
2366 {
2367 u_int i;
2368
2369 for (i = 0; i < np->num_queues; i++) {
2370 XN_RX_LOCK(&np->rxq[i]);
2371 XN_TX_LOCK(&np->txq[i]);
2372 }
2373 netfront_carrier_off(np);
2374 for (i = 0; i < np->num_queues; i++) {
2375 XN_RX_UNLOCK(&np->rxq[i]);
2376 XN_TX_UNLOCK(&np->txq[i]);
2377 }
2378
2379 for (i = 0; i < np->num_queues; i++) {
2380 disconnect_rxq(&np->rxq[i]);
2381 disconnect_txq(&np->txq[i]);
2382 }
2383 }
2384
2385 static int
2386 xn_ifmedia_upd(struct ifnet *ifp)
2387 {
2388
2389 return (0);
2390 }
2391
2392 static void
2393 xn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
2394 {
2395
2396 ifmr->ifm_status = IFM_AVALID|IFM_ACTIVE;
2397 ifmr->ifm_active = IFM_ETHER|IFM_MANUAL;
2398 }
2399
2400 /* ** Driver registration ** */
2401 static device_method_t netfront_methods[] = {
2402 /* Device interface */
2403 DEVMETHOD(device_probe, netfront_probe),
2404 DEVMETHOD(device_attach, netfront_attach),
2405 DEVMETHOD(device_detach, netfront_detach),
2406 DEVMETHOD(device_shutdown, bus_generic_shutdown),
2407 DEVMETHOD(device_suspend, netfront_suspend),
2408 DEVMETHOD(device_resume, netfront_resume),
2409
2410 /* Xenbus interface */
2411 DEVMETHOD(xenbus_otherend_changed, netfront_backend_changed),
2412
2413 DEVMETHOD_END
2414 };
2415
2416 static driver_t netfront_driver = {
2417 "xn",
2418 netfront_methods,
2419 sizeof(struct netfront_info),
2420 };
2421
2422 DRIVER_MODULE(xe, xenbusb_front, netfront_driver, NULL, NULL);
Cache object: 1c63616450a0ca6f7a56b92336bb12e9
|