Loading...
1/*
2 * Virtual network driver for conversing with remote driver backends.
3 *
4 * Copyright (c) 2002-2005, K A Fraser
5 * Copyright (c) 2005, XenSource Ltd
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License version 2
9 * as published by the Free Software Foundation; or, when distributed
10 * separately from the Linux kernel or incorporated into other
11 * software packages, subject to the following license:
12 *
13 * Permission is hereby granted, free of charge, to any person obtaining a copy
14 * of this source file (the "Software"), to deal in the Software without
15 * restriction, including without limitation the rights to use, copy, modify,
16 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
17 * and to permit persons to whom the Software is furnished to do so, subject to
18 * the following conditions:
19 *
20 * The above copyright notice and this permission notice shall be included in
21 * all copies or substantial portions of the Software.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
26 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
27 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
28 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
29 * IN THE SOFTWARE.
30 */
31
32#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
33
34#include <linux/module.h>
35#include <linux/kernel.h>
36#include <linux/netdevice.h>
37#include <linux/etherdevice.h>
38#include <linux/skbuff.h>
39#include <linux/ethtool.h>
40#include <linux/if_ether.h>
41#include <net/tcp.h>
42#include <linux/udp.h>
43#include <linux/moduleparam.h>
44#include <linux/mm.h>
45#include <linux/slab.h>
46#include <net/ip.h>
47#include <linux/bpf.h>
48#include <net/page_pool/types.h>
49#include <linux/bpf_trace.h>
50
51#include <xen/xen.h>
52#include <xen/xenbus.h>
53#include <xen/events.h>
54#include <xen/page.h>
55#include <xen/platform_pci.h>
56#include <xen/grant_table.h>
57
58#include <xen/interface/io/netif.h>
59#include <xen/interface/memory.h>
60#include <xen/interface/grant_table.h>
61
62/* Module parameters */
63#define MAX_QUEUES_DEFAULT 8
64static unsigned int xennet_max_queues;
65module_param_named(max_queues, xennet_max_queues, uint, 0644);
66MODULE_PARM_DESC(max_queues,
67 "Maximum number of queues per virtual interface");
68
69static bool __read_mostly xennet_trusted = true;
70module_param_named(trusted, xennet_trusted, bool, 0644);
71MODULE_PARM_DESC(trusted, "Is the backend trusted");
72
73#define XENNET_TIMEOUT (5 * HZ)
74
75static const struct ethtool_ops xennet_ethtool_ops;
76
77struct netfront_cb {
78 int pull_to;
79};
80
81#define NETFRONT_SKB_CB(skb) ((struct netfront_cb *)((skb)->cb))
82
83#define RX_COPY_THRESHOLD 256
84
85#define NET_TX_RING_SIZE __CONST_RING_SIZE(xen_netif_tx, XEN_PAGE_SIZE)
86#define NET_RX_RING_SIZE __CONST_RING_SIZE(xen_netif_rx, XEN_PAGE_SIZE)
87
88/* Minimum number of Rx slots (includes slot for GSO metadata). */
89#define NET_RX_SLOTS_MIN (XEN_NETIF_NR_SLOTS_MIN + 1)
90
91/* Queue name is interface name with "-qNNN" appended */
92#define QUEUE_NAME_SIZE (IFNAMSIZ + 6)
93
94/* IRQ name is queue name with "-tx" or "-rx" appended */
95#define IRQ_NAME_SIZE (QUEUE_NAME_SIZE + 3)
96
97static DECLARE_WAIT_QUEUE_HEAD(module_wq);
98
99struct netfront_stats {
100 u64 packets;
101 u64 bytes;
102 struct u64_stats_sync syncp;
103};
104
105struct netfront_info;
106
107struct netfront_queue {
108 unsigned int id; /* Queue ID, 0-based */
109 char name[QUEUE_NAME_SIZE]; /* DEVNAME-qN */
110 struct netfront_info *info;
111
112 struct bpf_prog __rcu *xdp_prog;
113
114 struct napi_struct napi;
115
116 /* Split event channels support, tx_* == rx_* when using
117 * single event channel.
118 */
119 unsigned int tx_evtchn, rx_evtchn;
120 unsigned int tx_irq, rx_irq;
121 /* Only used when split event channels support is enabled */
122 char tx_irq_name[IRQ_NAME_SIZE]; /* DEVNAME-qN-tx */
123 char rx_irq_name[IRQ_NAME_SIZE]; /* DEVNAME-qN-rx */
124
125 spinlock_t tx_lock;
126 struct xen_netif_tx_front_ring tx;
127 int tx_ring_ref;
128
129 /*
130 * {tx,rx}_skbs store outstanding skbuffs. Free tx_skb entries
131 * are linked from tx_skb_freelist through tx_link.
132 */
133 struct sk_buff *tx_skbs[NET_TX_RING_SIZE];
134 unsigned short tx_link[NET_TX_RING_SIZE];
135#define TX_LINK_NONE 0xffff
136#define TX_PENDING 0xfffe
137 grant_ref_t gref_tx_head;
138 grant_ref_t grant_tx_ref[NET_TX_RING_SIZE];
139 struct page *grant_tx_page[NET_TX_RING_SIZE];
140 unsigned tx_skb_freelist;
141 unsigned int tx_pend_queue;
142
143 spinlock_t rx_lock ____cacheline_aligned_in_smp;
144 struct xen_netif_rx_front_ring rx;
145 int rx_ring_ref;
146
147 struct timer_list rx_refill_timer;
148
149 struct sk_buff *rx_skbs[NET_RX_RING_SIZE];
150 grant_ref_t gref_rx_head;
151 grant_ref_t grant_rx_ref[NET_RX_RING_SIZE];
152
153 unsigned int rx_rsp_unconsumed;
154 spinlock_t rx_cons_lock;
155
156 struct page_pool *page_pool;
157 struct xdp_rxq_info xdp_rxq;
158};
159
160struct netfront_info {
161 struct list_head list;
162 struct net_device *netdev;
163
164 struct xenbus_device *xbdev;
165
166 /* Multi-queue support */
167 struct netfront_queue *queues;
168
169 /* Statistics */
170 struct netfront_stats __percpu *rx_stats;
171 struct netfront_stats __percpu *tx_stats;
172
173 /* XDP state */
174 bool netback_has_xdp_headroom;
175 bool netfront_xdp_enabled;
176
177 /* Is device behaving sane? */
178 bool broken;
179
180 /* Should skbs be bounced into a zeroed buffer? */
181 bool bounce;
182
183 atomic_t rx_gso_checksum_fixup;
184};
185
186struct netfront_rx_info {
187 struct xen_netif_rx_response rx;
188 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
189};
190
191/*
192 * Access macros for acquiring freeing slots in tx_skbs[].
193 */
194
195static void add_id_to_list(unsigned *head, unsigned short *list,
196 unsigned short id)
197{
198 list[id] = *head;
199 *head = id;
200}
201
202static unsigned short get_id_from_list(unsigned *head, unsigned short *list)
203{
204 unsigned int id = *head;
205
206 if (id != TX_LINK_NONE) {
207 *head = list[id];
208 list[id] = TX_LINK_NONE;
209 }
210 return id;
211}
212
213static int xennet_rxidx(RING_IDX idx)
214{
215 return idx & (NET_RX_RING_SIZE - 1);
216}
217
218static struct sk_buff *xennet_get_rx_skb(struct netfront_queue *queue,
219 RING_IDX ri)
220{
221 int i = xennet_rxidx(ri);
222 struct sk_buff *skb = queue->rx_skbs[i];
223 queue->rx_skbs[i] = NULL;
224 return skb;
225}
226
227static grant_ref_t xennet_get_rx_ref(struct netfront_queue *queue,
228 RING_IDX ri)
229{
230 int i = xennet_rxidx(ri);
231 grant_ref_t ref = queue->grant_rx_ref[i];
232 queue->grant_rx_ref[i] = INVALID_GRANT_REF;
233 return ref;
234}
235
236#ifdef CONFIG_SYSFS
237static const struct attribute_group xennet_dev_group;
238#endif
239
240static bool xennet_can_sg(struct net_device *dev)
241{
242 return dev->features & NETIF_F_SG;
243}
244
245
246static void rx_refill_timeout(struct timer_list *t)
247{
248 struct netfront_queue *queue = from_timer(queue, t, rx_refill_timer);
249 napi_schedule(&queue->napi);
250}
251
252static int netfront_tx_slot_available(struct netfront_queue *queue)
253{
254 return (queue->tx.req_prod_pvt - queue->tx.rsp_cons) <
255 (NET_TX_RING_SIZE - XEN_NETIF_NR_SLOTS_MIN - 1);
256}
257
258static void xennet_maybe_wake_tx(struct netfront_queue *queue)
259{
260 struct net_device *dev = queue->info->netdev;
261 struct netdev_queue *dev_queue = netdev_get_tx_queue(dev, queue->id);
262
263 if (unlikely(netif_tx_queue_stopped(dev_queue)) &&
264 netfront_tx_slot_available(queue) &&
265 likely(netif_running(dev)))
266 netif_tx_wake_queue(netdev_get_tx_queue(dev, queue->id));
267}
268
269
270static struct sk_buff *xennet_alloc_one_rx_buffer(struct netfront_queue *queue)
271{
272 struct sk_buff *skb;
273 struct page *page;
274
275 skb = __netdev_alloc_skb(queue->info->netdev,
276 RX_COPY_THRESHOLD + NET_IP_ALIGN,
277 GFP_ATOMIC | __GFP_NOWARN);
278 if (unlikely(!skb))
279 return NULL;
280
281 page = page_pool_alloc_pages(queue->page_pool,
282 GFP_ATOMIC | __GFP_NOWARN | __GFP_ZERO);
283 if (unlikely(!page)) {
284 kfree_skb(skb);
285 return NULL;
286 }
287 skb_add_rx_frag(skb, 0, page, 0, 0, PAGE_SIZE);
288
289 /* Align ip header to a 16 bytes boundary */
290 skb_reserve(skb, NET_IP_ALIGN);
291 skb->dev = queue->info->netdev;
292
293 return skb;
294}
295
296
297static void xennet_alloc_rx_buffers(struct netfront_queue *queue)
298{
299 RING_IDX req_prod = queue->rx.req_prod_pvt;
300 int notify;
301 int err = 0;
302
303 if (unlikely(!netif_carrier_ok(queue->info->netdev)))
304 return;
305
306 for (req_prod = queue->rx.req_prod_pvt;
307 req_prod - queue->rx.rsp_cons < NET_RX_RING_SIZE;
308 req_prod++) {
309 struct sk_buff *skb;
310 unsigned short id;
311 grant_ref_t ref;
312 struct page *page;
313 struct xen_netif_rx_request *req;
314
315 skb = xennet_alloc_one_rx_buffer(queue);
316 if (!skb) {
317 err = -ENOMEM;
318 break;
319 }
320
321 id = xennet_rxidx(req_prod);
322
323 BUG_ON(queue->rx_skbs[id]);
324 queue->rx_skbs[id] = skb;
325
326 ref = gnttab_claim_grant_reference(&queue->gref_rx_head);
327 WARN_ON_ONCE(IS_ERR_VALUE((unsigned long)(int)ref));
328 queue->grant_rx_ref[id] = ref;
329
330 page = skb_frag_page(&skb_shinfo(skb)->frags[0]);
331
332 req = RING_GET_REQUEST(&queue->rx, req_prod);
333 gnttab_page_grant_foreign_access_ref_one(ref,
334 queue->info->xbdev->otherend_id,
335 page,
336 0);
337 req->id = id;
338 req->gref = ref;
339 }
340
341 queue->rx.req_prod_pvt = req_prod;
342
343 /* Try again later if there are not enough requests or skb allocation
344 * failed.
345 * Enough requests is quantified as the sum of newly created slots and
346 * the unconsumed slots at the backend.
347 */
348 if (req_prod - queue->rx.rsp_cons < NET_RX_SLOTS_MIN ||
349 unlikely(err)) {
350 mod_timer(&queue->rx_refill_timer, jiffies + (HZ/10));
351 return;
352 }
353
354 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&queue->rx, notify);
355 if (notify)
356 notify_remote_via_irq(queue->rx_irq);
357}
358
359static int xennet_open(struct net_device *dev)
360{
361 struct netfront_info *np = netdev_priv(dev);
362 unsigned int num_queues = dev->real_num_tx_queues;
363 unsigned int i = 0;
364 struct netfront_queue *queue = NULL;
365
366 if (!np->queues || np->broken)
367 return -ENODEV;
368
369 for (i = 0; i < num_queues; ++i) {
370 queue = &np->queues[i];
371 napi_enable(&queue->napi);
372
373 spin_lock_bh(&queue->rx_lock);
374 if (netif_carrier_ok(dev)) {
375 xennet_alloc_rx_buffers(queue);
376 queue->rx.sring->rsp_event = queue->rx.rsp_cons + 1;
377 if (RING_HAS_UNCONSUMED_RESPONSES(&queue->rx))
378 napi_schedule(&queue->napi);
379 }
380 spin_unlock_bh(&queue->rx_lock);
381 }
382
383 netif_tx_start_all_queues(dev);
384
385 return 0;
386}
387
388static bool xennet_tx_buf_gc(struct netfront_queue *queue)
389{
390 RING_IDX cons, prod;
391 unsigned short id;
392 struct sk_buff *skb;
393 bool more_to_do;
394 bool work_done = false;
395 const struct device *dev = &queue->info->netdev->dev;
396
397 BUG_ON(!netif_carrier_ok(queue->info->netdev));
398
399 do {
400 prod = queue->tx.sring->rsp_prod;
401 if (RING_RESPONSE_PROD_OVERFLOW(&queue->tx, prod)) {
402 dev_alert(dev, "Illegal number of responses %u\n",
403 prod - queue->tx.rsp_cons);
404 goto err;
405 }
406 rmb(); /* Ensure we see responses up to 'rp'. */
407
408 for (cons = queue->tx.rsp_cons; cons != prod; cons++) {
409 struct xen_netif_tx_response txrsp;
410
411 work_done = true;
412
413 RING_COPY_RESPONSE(&queue->tx, cons, &txrsp);
414 if (txrsp.status == XEN_NETIF_RSP_NULL)
415 continue;
416
417 id = txrsp.id;
418 if (id >= RING_SIZE(&queue->tx)) {
419 dev_alert(dev,
420 "Response has incorrect id (%u)\n",
421 id);
422 goto err;
423 }
424 if (queue->tx_link[id] != TX_PENDING) {
425 dev_alert(dev,
426 "Response for inactive request\n");
427 goto err;
428 }
429
430 queue->tx_link[id] = TX_LINK_NONE;
431 skb = queue->tx_skbs[id];
432 queue->tx_skbs[id] = NULL;
433 if (unlikely(!gnttab_end_foreign_access_ref(
434 queue->grant_tx_ref[id]))) {
435 dev_alert(dev,
436 "Grant still in use by backend domain\n");
437 goto err;
438 }
439 gnttab_release_grant_reference(
440 &queue->gref_tx_head, queue->grant_tx_ref[id]);
441 queue->grant_tx_ref[id] = INVALID_GRANT_REF;
442 queue->grant_tx_page[id] = NULL;
443 add_id_to_list(&queue->tx_skb_freelist, queue->tx_link, id);
444 dev_kfree_skb_irq(skb);
445 }
446
447 queue->tx.rsp_cons = prod;
448
449 RING_FINAL_CHECK_FOR_RESPONSES(&queue->tx, more_to_do);
450 } while (more_to_do);
451
452 xennet_maybe_wake_tx(queue);
453
454 return work_done;
455
456 err:
457 queue->info->broken = true;
458 dev_alert(dev, "Disabled for further use\n");
459
460 return work_done;
461}
462
463struct xennet_gnttab_make_txreq {
464 struct netfront_queue *queue;
465 struct sk_buff *skb;
466 struct page *page;
467 struct xen_netif_tx_request *tx; /* Last request on ring page */
468 struct xen_netif_tx_request tx_local; /* Last request local copy*/
469 unsigned int size;
470};
471
472static void xennet_tx_setup_grant(unsigned long gfn, unsigned int offset,
473 unsigned int len, void *data)
474{
475 struct xennet_gnttab_make_txreq *info = data;
476 unsigned int id;
477 struct xen_netif_tx_request *tx;
478 grant_ref_t ref;
479 /* convenient aliases */
480 struct page *page = info->page;
481 struct netfront_queue *queue = info->queue;
482 struct sk_buff *skb = info->skb;
483
484 id = get_id_from_list(&queue->tx_skb_freelist, queue->tx_link);
485 tx = RING_GET_REQUEST(&queue->tx, queue->tx.req_prod_pvt++);
486 ref = gnttab_claim_grant_reference(&queue->gref_tx_head);
487 WARN_ON_ONCE(IS_ERR_VALUE((unsigned long)(int)ref));
488
489 gnttab_grant_foreign_access_ref(ref, queue->info->xbdev->otherend_id,
490 gfn, GNTMAP_readonly);
491
492 queue->tx_skbs[id] = skb;
493 queue->grant_tx_page[id] = page;
494 queue->grant_tx_ref[id] = ref;
495
496 info->tx_local.id = id;
497 info->tx_local.gref = ref;
498 info->tx_local.offset = offset;
499 info->tx_local.size = len;
500 info->tx_local.flags = 0;
501
502 *tx = info->tx_local;
503
504 /*
505 * Put the request in the pending queue, it will be set to be pending
506 * when the producer index is about to be raised.
507 */
508 add_id_to_list(&queue->tx_pend_queue, queue->tx_link, id);
509
510 info->tx = tx;
511 info->size += info->tx_local.size;
512}
513
514static struct xen_netif_tx_request *xennet_make_first_txreq(
515 struct xennet_gnttab_make_txreq *info,
516 unsigned int offset, unsigned int len)
517{
518 info->size = 0;
519
520 gnttab_for_one_grant(info->page, offset, len, xennet_tx_setup_grant, info);
521
522 return info->tx;
523}
524
525static void xennet_make_one_txreq(unsigned long gfn, unsigned int offset,
526 unsigned int len, void *data)
527{
528 struct xennet_gnttab_make_txreq *info = data;
529
530 info->tx->flags |= XEN_NETTXF_more_data;
531 skb_get(info->skb);
532 xennet_tx_setup_grant(gfn, offset, len, data);
533}
534
535static void xennet_make_txreqs(
536 struct xennet_gnttab_make_txreq *info,
537 struct page *page,
538 unsigned int offset, unsigned int len)
539{
540 /* Skip unused frames from start of page */
541 page += offset >> PAGE_SHIFT;
542 offset &= ~PAGE_MASK;
543
544 while (len) {
545 info->page = page;
546 info->size = 0;
547
548 gnttab_foreach_grant_in_range(page, offset, len,
549 xennet_make_one_txreq,
550 info);
551
552 page++;
553 offset = 0;
554 len -= info->size;
555 }
556}
557
558/*
559 * Count how many ring slots are required to send this skb. Each frag
560 * might be a compound page.
561 */
562static int xennet_count_skb_slots(struct sk_buff *skb)
563{
564 int i, frags = skb_shinfo(skb)->nr_frags;
565 int slots;
566
567 slots = gnttab_count_grant(offset_in_page(skb->data),
568 skb_headlen(skb));
569
570 for (i = 0; i < frags; i++) {
571 skb_frag_t *frag = skb_shinfo(skb)->frags + i;
572 unsigned long size = skb_frag_size(frag);
573 unsigned long offset = skb_frag_off(frag);
574
575 /* Skip unused frames from start of page */
576 offset &= ~PAGE_MASK;
577
578 slots += gnttab_count_grant(offset, size);
579 }
580
581 return slots;
582}
583
584static u16 xennet_select_queue(struct net_device *dev, struct sk_buff *skb,
585 struct net_device *sb_dev)
586{
587 unsigned int num_queues = dev->real_num_tx_queues;
588 u32 hash;
589 u16 queue_idx;
590
591 /* First, check if there is only one queue */
592 if (num_queues == 1) {
593 queue_idx = 0;
594 } else {
595 hash = skb_get_hash(skb);
596 queue_idx = hash % num_queues;
597 }
598
599 return queue_idx;
600}
601
602static void xennet_mark_tx_pending(struct netfront_queue *queue)
603{
604 unsigned int i;
605
606 while ((i = get_id_from_list(&queue->tx_pend_queue, queue->tx_link)) !=
607 TX_LINK_NONE)
608 queue->tx_link[i] = TX_PENDING;
609}
610
611static int xennet_xdp_xmit_one(struct net_device *dev,
612 struct netfront_queue *queue,
613 struct xdp_frame *xdpf)
614{
615 struct netfront_info *np = netdev_priv(dev);
616 struct netfront_stats *tx_stats = this_cpu_ptr(np->tx_stats);
617 struct xennet_gnttab_make_txreq info = {
618 .queue = queue,
619 .skb = NULL,
620 .page = virt_to_page(xdpf->data),
621 };
622 int notify;
623
624 xennet_make_first_txreq(&info,
625 offset_in_page(xdpf->data),
626 xdpf->len);
627
628 xennet_mark_tx_pending(queue);
629
630 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&queue->tx, notify);
631 if (notify)
632 notify_remote_via_irq(queue->tx_irq);
633
634 u64_stats_update_begin(&tx_stats->syncp);
635 tx_stats->bytes += xdpf->len;
636 tx_stats->packets++;
637 u64_stats_update_end(&tx_stats->syncp);
638
639 xennet_tx_buf_gc(queue);
640
641 return 0;
642}
643
644static int xennet_xdp_xmit(struct net_device *dev, int n,
645 struct xdp_frame **frames, u32 flags)
646{
647 unsigned int num_queues = dev->real_num_tx_queues;
648 struct netfront_info *np = netdev_priv(dev);
649 struct netfront_queue *queue = NULL;
650 unsigned long irq_flags;
651 int nxmit = 0;
652 int i;
653
654 if (unlikely(np->broken))
655 return -ENODEV;
656 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
657 return -EINVAL;
658
659 queue = &np->queues[smp_processor_id() % num_queues];
660
661 spin_lock_irqsave(&queue->tx_lock, irq_flags);
662 for (i = 0; i < n; i++) {
663 struct xdp_frame *xdpf = frames[i];
664
665 if (!xdpf)
666 continue;
667 if (xennet_xdp_xmit_one(dev, queue, xdpf))
668 break;
669 nxmit++;
670 }
671 spin_unlock_irqrestore(&queue->tx_lock, irq_flags);
672
673 return nxmit;
674}
675
676static struct sk_buff *bounce_skb(const struct sk_buff *skb)
677{
678 unsigned int headerlen = skb_headroom(skb);
679 /* Align size to allocate full pages and avoid contiguous data leaks */
680 unsigned int size = ALIGN(skb_end_offset(skb) + skb->data_len,
681 XEN_PAGE_SIZE);
682 struct sk_buff *n = alloc_skb(size, GFP_ATOMIC | __GFP_ZERO);
683
684 if (!n)
685 return NULL;
686
687 if (!IS_ALIGNED((uintptr_t)n->head, XEN_PAGE_SIZE)) {
688 WARN_ONCE(1, "misaligned skb allocated\n");
689 kfree_skb(n);
690 return NULL;
691 }
692
693 /* Set the data pointer */
694 skb_reserve(n, headerlen);
695 /* Set the tail pointer and length */
696 skb_put(n, skb->len);
697
698 BUG_ON(skb_copy_bits(skb, -headerlen, n->head, headerlen + skb->len));
699
700 skb_copy_header(n, skb);
701 return n;
702}
703
704#define MAX_XEN_SKB_FRAGS (65536 / XEN_PAGE_SIZE + 1)
705
706static netdev_tx_t xennet_start_xmit(struct sk_buff *skb, struct net_device *dev)
707{
708 struct netfront_info *np = netdev_priv(dev);
709 struct netfront_stats *tx_stats = this_cpu_ptr(np->tx_stats);
710 struct xen_netif_tx_request *first_tx;
711 unsigned int i;
712 int notify;
713 int slots;
714 struct page *page;
715 unsigned int offset;
716 unsigned int len;
717 unsigned long flags;
718 struct netfront_queue *queue = NULL;
719 struct xennet_gnttab_make_txreq info = { };
720 unsigned int num_queues = dev->real_num_tx_queues;
721 u16 queue_index;
722 struct sk_buff *nskb;
723
724 /* Drop the packet if no queues are set up */
725 if (num_queues < 1)
726 goto drop;
727 if (unlikely(np->broken))
728 goto drop;
729 /* Determine which queue to transmit this SKB on */
730 queue_index = skb_get_queue_mapping(skb);
731 queue = &np->queues[queue_index];
732
733 /* If skb->len is too big for wire format, drop skb and alert
734 * user about misconfiguration.
735 */
736 if (unlikely(skb->len > XEN_NETIF_MAX_TX_SIZE)) {
737 net_alert_ratelimited(
738 "xennet: skb->len = %u, too big for wire format\n",
739 skb->len);
740 goto drop;
741 }
742
743 slots = xennet_count_skb_slots(skb);
744 if (unlikely(slots > MAX_XEN_SKB_FRAGS + 1)) {
745 net_dbg_ratelimited("xennet: skb rides the rocket: %d slots, %d bytes\n",
746 slots, skb->len);
747 if (skb_linearize(skb))
748 goto drop;
749 }
750
751 page = virt_to_page(skb->data);
752 offset = offset_in_page(skb->data);
753
754 /* The first req should be at least ETH_HLEN size or the packet will be
755 * dropped by netback.
756 *
757 * If the backend is not trusted bounce all data to zeroed pages to
758 * avoid exposing contiguous data on the granted page not belonging to
759 * the skb.
760 */
761 if (np->bounce || unlikely(PAGE_SIZE - offset < ETH_HLEN)) {
762 nskb = bounce_skb(skb);
763 if (!nskb)
764 goto drop;
765 dev_consume_skb_any(skb);
766 skb = nskb;
767 page = virt_to_page(skb->data);
768 offset = offset_in_page(skb->data);
769 }
770
771 len = skb_headlen(skb);
772
773 spin_lock_irqsave(&queue->tx_lock, flags);
774
775 if (unlikely(!netif_carrier_ok(dev) ||
776 (slots > 1 && !xennet_can_sg(dev)) ||
777 netif_needs_gso(skb, netif_skb_features(skb)))) {
778 spin_unlock_irqrestore(&queue->tx_lock, flags);
779 goto drop;
780 }
781
782 /* First request for the linear area. */
783 info.queue = queue;
784 info.skb = skb;
785 info.page = page;
786 first_tx = xennet_make_first_txreq(&info, offset, len);
787 offset += info.tx_local.size;
788 if (offset == PAGE_SIZE) {
789 page++;
790 offset = 0;
791 }
792 len -= info.tx_local.size;
793
794 if (skb->ip_summed == CHECKSUM_PARTIAL)
795 /* local packet? */
796 first_tx->flags |= XEN_NETTXF_csum_blank |
797 XEN_NETTXF_data_validated;
798 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
799 /* remote but checksummed. */
800 first_tx->flags |= XEN_NETTXF_data_validated;
801
802 /* Optional extra info after the first request. */
803 if (skb_shinfo(skb)->gso_size) {
804 struct xen_netif_extra_info *gso;
805
806 gso = (struct xen_netif_extra_info *)
807 RING_GET_REQUEST(&queue->tx, queue->tx.req_prod_pvt++);
808
809 first_tx->flags |= XEN_NETTXF_extra_info;
810
811 gso->u.gso.size = skb_shinfo(skb)->gso_size;
812 gso->u.gso.type = (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) ?
813 XEN_NETIF_GSO_TYPE_TCPV6 :
814 XEN_NETIF_GSO_TYPE_TCPV4;
815 gso->u.gso.pad = 0;
816 gso->u.gso.features = 0;
817
818 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
819 gso->flags = 0;
820 }
821
822 /* Requests for the rest of the linear area. */
823 xennet_make_txreqs(&info, page, offset, len);
824
825 /* Requests for all the frags. */
826 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
827 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
828 xennet_make_txreqs(&info, skb_frag_page(frag),
829 skb_frag_off(frag),
830 skb_frag_size(frag));
831 }
832
833 /* First request has the packet length. */
834 first_tx->size = skb->len;
835
836 /* timestamp packet in software */
837 skb_tx_timestamp(skb);
838
839 xennet_mark_tx_pending(queue);
840
841 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&queue->tx, notify);
842 if (notify)
843 notify_remote_via_irq(queue->tx_irq);
844
845 u64_stats_update_begin(&tx_stats->syncp);
846 tx_stats->bytes += skb->len;
847 tx_stats->packets++;
848 u64_stats_update_end(&tx_stats->syncp);
849
850 /* Note: It is not safe to access skb after xennet_tx_buf_gc()! */
851 xennet_tx_buf_gc(queue);
852
853 if (!netfront_tx_slot_available(queue))
854 netif_tx_stop_queue(netdev_get_tx_queue(dev, queue->id));
855
856 spin_unlock_irqrestore(&queue->tx_lock, flags);
857
858 return NETDEV_TX_OK;
859
860 drop:
861 dev->stats.tx_dropped++;
862 dev_kfree_skb_any(skb);
863 return NETDEV_TX_OK;
864}
865
866static int xennet_close(struct net_device *dev)
867{
868 struct netfront_info *np = netdev_priv(dev);
869 unsigned int num_queues = dev->real_num_tx_queues;
870 unsigned int i;
871 struct netfront_queue *queue;
872 netif_tx_stop_all_queues(np->netdev);
873 for (i = 0; i < num_queues; ++i) {
874 queue = &np->queues[i];
875 napi_disable(&queue->napi);
876 }
877 return 0;
878}
879
880static void xennet_destroy_queues(struct netfront_info *info)
881{
882 unsigned int i;
883
884 for (i = 0; i < info->netdev->real_num_tx_queues; i++) {
885 struct netfront_queue *queue = &info->queues[i];
886
887 if (netif_running(info->netdev))
888 napi_disable(&queue->napi);
889 netif_napi_del(&queue->napi);
890 }
891
892 kfree(info->queues);
893 info->queues = NULL;
894}
895
896static void xennet_uninit(struct net_device *dev)
897{
898 struct netfront_info *np = netdev_priv(dev);
899 xennet_destroy_queues(np);
900}
901
902static void xennet_set_rx_rsp_cons(struct netfront_queue *queue, RING_IDX val)
903{
904 unsigned long flags;
905
906 spin_lock_irqsave(&queue->rx_cons_lock, flags);
907 queue->rx.rsp_cons = val;
908 queue->rx_rsp_unconsumed = XEN_RING_NR_UNCONSUMED_RESPONSES(&queue->rx);
909 spin_unlock_irqrestore(&queue->rx_cons_lock, flags);
910}
911
912static void xennet_move_rx_slot(struct netfront_queue *queue, struct sk_buff *skb,
913 grant_ref_t ref)
914{
915 int new = xennet_rxidx(queue->rx.req_prod_pvt);
916
917 BUG_ON(queue->rx_skbs[new]);
918 queue->rx_skbs[new] = skb;
919 queue->grant_rx_ref[new] = ref;
920 RING_GET_REQUEST(&queue->rx, queue->rx.req_prod_pvt)->id = new;
921 RING_GET_REQUEST(&queue->rx, queue->rx.req_prod_pvt)->gref = ref;
922 queue->rx.req_prod_pvt++;
923}
924
925static int xennet_get_extras(struct netfront_queue *queue,
926 struct xen_netif_extra_info *extras,
927 RING_IDX rp)
928
929{
930 struct xen_netif_extra_info extra;
931 struct device *dev = &queue->info->netdev->dev;
932 RING_IDX cons = queue->rx.rsp_cons;
933 int err = 0;
934
935 do {
936 struct sk_buff *skb;
937 grant_ref_t ref;
938
939 if (unlikely(cons + 1 == rp)) {
940 if (net_ratelimit())
941 dev_warn(dev, "Missing extra info\n");
942 err = -EBADR;
943 break;
944 }
945
946 RING_COPY_RESPONSE(&queue->rx, ++cons, &extra);
947
948 if (unlikely(!extra.type ||
949 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
950 if (net_ratelimit())
951 dev_warn(dev, "Invalid extra type: %d\n",
952 extra.type);
953 err = -EINVAL;
954 } else {
955 extras[extra.type - 1] = extra;
956 }
957
958 skb = xennet_get_rx_skb(queue, cons);
959 ref = xennet_get_rx_ref(queue, cons);
960 xennet_move_rx_slot(queue, skb, ref);
961 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
962
963 xennet_set_rx_rsp_cons(queue, cons);
964 return err;
965}
966
967static u32 xennet_run_xdp(struct netfront_queue *queue, struct page *pdata,
968 struct xen_netif_rx_response *rx, struct bpf_prog *prog,
969 struct xdp_buff *xdp, bool *need_xdp_flush)
970{
971 struct xdp_frame *xdpf;
972 u32 len = rx->status;
973 u32 act;
974 int err;
975
976 xdp_init_buff(xdp, XEN_PAGE_SIZE - XDP_PACKET_HEADROOM,
977 &queue->xdp_rxq);
978 xdp_prepare_buff(xdp, page_address(pdata), XDP_PACKET_HEADROOM,
979 len, false);
980
981 act = bpf_prog_run_xdp(prog, xdp);
982 switch (act) {
983 case XDP_TX:
984 get_page(pdata);
985 xdpf = xdp_convert_buff_to_frame(xdp);
986 err = xennet_xdp_xmit(queue->info->netdev, 1, &xdpf, 0);
987 if (unlikely(!err))
988 xdp_return_frame_rx_napi(xdpf);
989 else if (unlikely(err < 0))
990 trace_xdp_exception(queue->info->netdev, prog, act);
991 break;
992 case XDP_REDIRECT:
993 get_page(pdata);
994 err = xdp_do_redirect(queue->info->netdev, xdp, prog);
995 *need_xdp_flush = true;
996 if (unlikely(err))
997 trace_xdp_exception(queue->info->netdev, prog, act);
998 break;
999 case XDP_PASS:
1000 case XDP_DROP:
1001 break;
1002
1003 case XDP_ABORTED:
1004 trace_xdp_exception(queue->info->netdev, prog, act);
1005 break;
1006
1007 default:
1008 bpf_warn_invalid_xdp_action(queue->info->netdev, prog, act);
1009 }
1010
1011 return act;
1012}
1013
1014static int xennet_get_responses(struct netfront_queue *queue,
1015 struct netfront_rx_info *rinfo, RING_IDX rp,
1016 struct sk_buff_head *list,
1017 bool *need_xdp_flush)
1018{
1019 struct xen_netif_rx_response *rx = &rinfo->rx, rx_local;
1020 int max = XEN_NETIF_NR_SLOTS_MIN + (rx->status <= RX_COPY_THRESHOLD);
1021 RING_IDX cons = queue->rx.rsp_cons;
1022 struct sk_buff *skb = xennet_get_rx_skb(queue, cons);
1023 struct xen_netif_extra_info *extras = rinfo->extras;
1024 grant_ref_t ref = xennet_get_rx_ref(queue, cons);
1025 struct device *dev = &queue->info->netdev->dev;
1026 struct bpf_prog *xdp_prog;
1027 struct xdp_buff xdp;
1028 int slots = 1;
1029 int err = 0;
1030 u32 verdict;
1031
1032 if (rx->flags & XEN_NETRXF_extra_info) {
1033 err = xennet_get_extras(queue, extras, rp);
1034 if (!err) {
1035 if (extras[XEN_NETIF_EXTRA_TYPE_XDP - 1].type) {
1036 struct xen_netif_extra_info *xdp;
1037
1038 xdp = &extras[XEN_NETIF_EXTRA_TYPE_XDP - 1];
1039 rx->offset = xdp->u.xdp.headroom;
1040 }
1041 }
1042 cons = queue->rx.rsp_cons;
1043 }
1044
1045 for (;;) {
1046 /*
1047 * This definitely indicates a bug, either in this driver or in
1048 * the backend driver. In future this should flag the bad
1049 * situation to the system controller to reboot the backend.
1050 */
1051 if (ref == INVALID_GRANT_REF) {
1052 if (net_ratelimit())
1053 dev_warn(dev, "Bad rx response id %d.\n",
1054 rx->id);
1055 err = -EINVAL;
1056 goto next;
1057 }
1058
1059 if (unlikely(rx->status < 0 ||
1060 rx->offset + rx->status > XEN_PAGE_SIZE)) {
1061 if (net_ratelimit())
1062 dev_warn(dev, "rx->offset: %u, size: %d\n",
1063 rx->offset, rx->status);
1064 xennet_move_rx_slot(queue, skb, ref);
1065 err = -EINVAL;
1066 goto next;
1067 }
1068
1069 if (!gnttab_end_foreign_access_ref(ref)) {
1070 dev_alert(dev,
1071 "Grant still in use by backend domain\n");
1072 queue->info->broken = true;
1073 dev_alert(dev, "Disabled for further use\n");
1074 return -EINVAL;
1075 }
1076
1077 gnttab_release_grant_reference(&queue->gref_rx_head, ref);
1078
1079 rcu_read_lock();
1080 xdp_prog = rcu_dereference(queue->xdp_prog);
1081 if (xdp_prog) {
1082 if (!(rx->flags & XEN_NETRXF_more_data)) {
1083 /* currently only a single page contains data */
1084 verdict = xennet_run_xdp(queue,
1085 skb_frag_page(&skb_shinfo(skb)->frags[0]),
1086 rx, xdp_prog, &xdp, need_xdp_flush);
1087 if (verdict != XDP_PASS)
1088 err = -EINVAL;
1089 } else {
1090 /* drop the frame */
1091 err = -EINVAL;
1092 }
1093 }
1094 rcu_read_unlock();
1095
1096 __skb_queue_tail(list, skb);
1097
1098next:
1099 if (!(rx->flags & XEN_NETRXF_more_data))
1100 break;
1101
1102 if (cons + slots == rp) {
1103 if (net_ratelimit())
1104 dev_warn(dev, "Need more slots\n");
1105 err = -ENOENT;
1106 break;
1107 }
1108
1109 RING_COPY_RESPONSE(&queue->rx, cons + slots, &rx_local);
1110 rx = &rx_local;
1111 skb = xennet_get_rx_skb(queue, cons + slots);
1112 ref = xennet_get_rx_ref(queue, cons + slots);
1113 slots++;
1114 }
1115
1116 if (unlikely(slots > max)) {
1117 if (net_ratelimit())
1118 dev_warn(dev, "Too many slots\n");
1119 err = -E2BIG;
1120 }
1121
1122 if (unlikely(err))
1123 xennet_set_rx_rsp_cons(queue, cons + slots);
1124
1125 return err;
1126}
1127
1128static int xennet_set_skb_gso(struct sk_buff *skb,
1129 struct xen_netif_extra_info *gso)
1130{
1131 if (!gso->u.gso.size) {
1132 if (net_ratelimit())
1133 pr_warn("GSO size must not be zero\n");
1134 return -EINVAL;
1135 }
1136
1137 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4 &&
1138 gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV6) {
1139 if (net_ratelimit())
1140 pr_warn("Bad GSO type %d\n", gso->u.gso.type);
1141 return -EINVAL;
1142 }
1143
1144 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1145 skb_shinfo(skb)->gso_type =
1146 (gso->u.gso.type == XEN_NETIF_GSO_TYPE_TCPV4) ?
1147 SKB_GSO_TCPV4 :
1148 SKB_GSO_TCPV6;
1149
1150 /* Header must be checked, and gso_segs computed. */
1151 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1152 skb_shinfo(skb)->gso_segs = 0;
1153
1154 return 0;
1155}
1156
1157static int xennet_fill_frags(struct netfront_queue *queue,
1158 struct sk_buff *skb,
1159 struct sk_buff_head *list)
1160{
1161 RING_IDX cons = queue->rx.rsp_cons;
1162 struct sk_buff *nskb;
1163
1164 while ((nskb = __skb_dequeue(list))) {
1165 struct xen_netif_rx_response rx;
1166 skb_frag_t *nfrag = &skb_shinfo(nskb)->frags[0];
1167
1168 RING_COPY_RESPONSE(&queue->rx, ++cons, &rx);
1169
1170 if (skb_shinfo(skb)->nr_frags == MAX_SKB_FRAGS) {
1171 unsigned int pull_to = NETFRONT_SKB_CB(skb)->pull_to;
1172
1173 BUG_ON(pull_to < skb_headlen(skb));
1174 __pskb_pull_tail(skb, pull_to - skb_headlen(skb));
1175 }
1176 if (unlikely(skb_shinfo(skb)->nr_frags >= MAX_SKB_FRAGS)) {
1177 xennet_set_rx_rsp_cons(queue,
1178 ++cons + skb_queue_len(list));
1179 kfree_skb(nskb);
1180 return -ENOENT;
1181 }
1182
1183 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
1184 skb_frag_page(nfrag),
1185 rx.offset, rx.status, PAGE_SIZE);
1186
1187 skb_shinfo(nskb)->nr_frags = 0;
1188 kfree_skb(nskb);
1189 }
1190
1191 xennet_set_rx_rsp_cons(queue, cons);
1192
1193 return 0;
1194}
1195
1196static int checksum_setup(struct net_device *dev, struct sk_buff *skb)
1197{
1198 bool recalculate_partial_csum = false;
1199
1200 /*
1201 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1202 * peers can fail to set NETRXF_csum_blank when sending a GSO
1203 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1204 * recalculate the partial checksum.
1205 */
1206 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1207 struct netfront_info *np = netdev_priv(dev);
1208 atomic_inc(&np->rx_gso_checksum_fixup);
1209 skb->ip_summed = CHECKSUM_PARTIAL;
1210 recalculate_partial_csum = true;
1211 }
1212
1213 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1214 if (skb->ip_summed != CHECKSUM_PARTIAL)
1215 return 0;
1216
1217 return skb_checksum_setup(skb, recalculate_partial_csum);
1218}
1219
1220static int handle_incoming_queue(struct netfront_queue *queue,
1221 struct sk_buff_head *rxq)
1222{
1223 struct netfront_stats *rx_stats = this_cpu_ptr(queue->info->rx_stats);
1224 int packets_dropped = 0;
1225 struct sk_buff *skb;
1226
1227 while ((skb = __skb_dequeue(rxq)) != NULL) {
1228 int pull_to = NETFRONT_SKB_CB(skb)->pull_to;
1229
1230 if (pull_to > skb_headlen(skb))
1231 __pskb_pull_tail(skb, pull_to - skb_headlen(skb));
1232
1233 /* Ethernet work: Delayed to here as it peeks the header. */
1234 skb->protocol = eth_type_trans(skb, queue->info->netdev);
1235 skb_reset_network_header(skb);
1236
1237 if (checksum_setup(queue->info->netdev, skb)) {
1238 kfree_skb(skb);
1239 packets_dropped++;
1240 queue->info->netdev->stats.rx_errors++;
1241 continue;
1242 }
1243
1244 u64_stats_update_begin(&rx_stats->syncp);
1245 rx_stats->packets++;
1246 rx_stats->bytes += skb->len;
1247 u64_stats_update_end(&rx_stats->syncp);
1248
1249 /* Pass it up. */
1250 napi_gro_receive(&queue->napi, skb);
1251 }
1252
1253 return packets_dropped;
1254}
1255
1256static int xennet_poll(struct napi_struct *napi, int budget)
1257{
1258 struct netfront_queue *queue = container_of(napi, struct netfront_queue, napi);
1259 struct net_device *dev = queue->info->netdev;
1260 struct sk_buff *skb;
1261 struct netfront_rx_info rinfo;
1262 struct xen_netif_rx_response *rx = &rinfo.rx;
1263 struct xen_netif_extra_info *extras = rinfo.extras;
1264 RING_IDX i, rp;
1265 int work_done;
1266 struct sk_buff_head rxq;
1267 struct sk_buff_head errq;
1268 struct sk_buff_head tmpq;
1269 int err;
1270 bool need_xdp_flush = false;
1271
1272 spin_lock(&queue->rx_lock);
1273
1274 skb_queue_head_init(&rxq);
1275 skb_queue_head_init(&errq);
1276 skb_queue_head_init(&tmpq);
1277
1278 rp = queue->rx.sring->rsp_prod;
1279 if (RING_RESPONSE_PROD_OVERFLOW(&queue->rx, rp)) {
1280 dev_alert(&dev->dev, "Illegal number of responses %u\n",
1281 rp - queue->rx.rsp_cons);
1282 queue->info->broken = true;
1283 spin_unlock(&queue->rx_lock);
1284 return 0;
1285 }
1286 rmb(); /* Ensure we see queued responses up to 'rp'. */
1287
1288 i = queue->rx.rsp_cons;
1289 work_done = 0;
1290 while ((i != rp) && (work_done < budget)) {
1291 RING_COPY_RESPONSE(&queue->rx, i, rx);
1292 memset(extras, 0, sizeof(rinfo.extras));
1293
1294 err = xennet_get_responses(queue, &rinfo, rp, &tmpq,
1295 &need_xdp_flush);
1296
1297 if (unlikely(err)) {
1298 if (queue->info->broken) {
1299 spin_unlock(&queue->rx_lock);
1300 return 0;
1301 }
1302err:
1303 while ((skb = __skb_dequeue(&tmpq)))
1304 __skb_queue_tail(&errq, skb);
1305 dev->stats.rx_errors++;
1306 i = queue->rx.rsp_cons;
1307 continue;
1308 }
1309
1310 skb = __skb_dequeue(&tmpq);
1311
1312 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1313 struct xen_netif_extra_info *gso;
1314 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1315
1316 if (unlikely(xennet_set_skb_gso(skb, gso))) {
1317 __skb_queue_head(&tmpq, skb);
1318 xennet_set_rx_rsp_cons(queue,
1319 queue->rx.rsp_cons +
1320 skb_queue_len(&tmpq));
1321 goto err;
1322 }
1323 }
1324
1325 NETFRONT_SKB_CB(skb)->pull_to = rx->status;
1326 if (NETFRONT_SKB_CB(skb)->pull_to > RX_COPY_THRESHOLD)
1327 NETFRONT_SKB_CB(skb)->pull_to = RX_COPY_THRESHOLD;
1328
1329 skb_frag_off_set(&skb_shinfo(skb)->frags[0], rx->offset);
1330 skb_frag_size_set(&skb_shinfo(skb)->frags[0], rx->status);
1331 skb->data_len = rx->status;
1332 skb->len += rx->status;
1333
1334 if (unlikely(xennet_fill_frags(queue, skb, &tmpq)))
1335 goto err;
1336
1337 if (rx->flags & XEN_NETRXF_csum_blank)
1338 skb->ip_summed = CHECKSUM_PARTIAL;
1339 else if (rx->flags & XEN_NETRXF_data_validated)
1340 skb->ip_summed = CHECKSUM_UNNECESSARY;
1341
1342 __skb_queue_tail(&rxq, skb);
1343
1344 i = queue->rx.rsp_cons + 1;
1345 xennet_set_rx_rsp_cons(queue, i);
1346 work_done++;
1347 }
1348 if (need_xdp_flush)
1349 xdp_do_flush();
1350
1351 __skb_queue_purge(&errq);
1352
1353 work_done -= handle_incoming_queue(queue, &rxq);
1354
1355 xennet_alloc_rx_buffers(queue);
1356
1357 if (work_done < budget) {
1358 int more_to_do = 0;
1359
1360 napi_complete_done(napi, work_done);
1361
1362 RING_FINAL_CHECK_FOR_RESPONSES(&queue->rx, more_to_do);
1363 if (more_to_do)
1364 napi_schedule(napi);
1365 }
1366
1367 spin_unlock(&queue->rx_lock);
1368
1369 return work_done;
1370}
1371
1372static int xennet_change_mtu(struct net_device *dev, int mtu)
1373{
1374 int max = xennet_can_sg(dev) ? XEN_NETIF_MAX_TX_SIZE : ETH_DATA_LEN;
1375
1376 if (mtu > max)
1377 return -EINVAL;
1378 dev->mtu = mtu;
1379 return 0;
1380}
1381
1382static void xennet_get_stats64(struct net_device *dev,
1383 struct rtnl_link_stats64 *tot)
1384{
1385 struct netfront_info *np = netdev_priv(dev);
1386 int cpu;
1387
1388 for_each_possible_cpu(cpu) {
1389 struct netfront_stats *rx_stats = per_cpu_ptr(np->rx_stats, cpu);
1390 struct netfront_stats *tx_stats = per_cpu_ptr(np->tx_stats, cpu);
1391 u64 rx_packets, rx_bytes, tx_packets, tx_bytes;
1392 unsigned int start;
1393
1394 do {
1395 start = u64_stats_fetch_begin(&tx_stats->syncp);
1396 tx_packets = tx_stats->packets;
1397 tx_bytes = tx_stats->bytes;
1398 } while (u64_stats_fetch_retry(&tx_stats->syncp, start));
1399
1400 do {
1401 start = u64_stats_fetch_begin(&rx_stats->syncp);
1402 rx_packets = rx_stats->packets;
1403 rx_bytes = rx_stats->bytes;
1404 } while (u64_stats_fetch_retry(&rx_stats->syncp, start));
1405
1406 tot->rx_packets += rx_packets;
1407 tot->tx_packets += tx_packets;
1408 tot->rx_bytes += rx_bytes;
1409 tot->tx_bytes += tx_bytes;
1410 }
1411
1412 tot->rx_errors = dev->stats.rx_errors;
1413 tot->tx_dropped = dev->stats.tx_dropped;
1414}
1415
1416static void xennet_release_tx_bufs(struct netfront_queue *queue)
1417{
1418 struct sk_buff *skb;
1419 int i;
1420
1421 for (i = 0; i < NET_TX_RING_SIZE; i++) {
1422 /* Skip over entries which are actually freelist references */
1423 if (!queue->tx_skbs[i])
1424 continue;
1425
1426 skb = queue->tx_skbs[i];
1427 queue->tx_skbs[i] = NULL;
1428 get_page(queue->grant_tx_page[i]);
1429 gnttab_end_foreign_access(queue->grant_tx_ref[i],
1430 queue->grant_tx_page[i]);
1431 queue->grant_tx_page[i] = NULL;
1432 queue->grant_tx_ref[i] = INVALID_GRANT_REF;
1433 add_id_to_list(&queue->tx_skb_freelist, queue->tx_link, i);
1434 dev_kfree_skb_irq(skb);
1435 }
1436}
1437
1438static void xennet_release_rx_bufs(struct netfront_queue *queue)
1439{
1440 int id, ref;
1441
1442 spin_lock_bh(&queue->rx_lock);
1443
1444 for (id = 0; id < NET_RX_RING_SIZE; id++) {
1445 struct sk_buff *skb;
1446 struct page *page;
1447
1448 skb = queue->rx_skbs[id];
1449 if (!skb)
1450 continue;
1451
1452 ref = queue->grant_rx_ref[id];
1453 if (ref == INVALID_GRANT_REF)
1454 continue;
1455
1456 page = skb_frag_page(&skb_shinfo(skb)->frags[0]);
1457
1458 /* gnttab_end_foreign_access() needs a page ref until
1459 * foreign access is ended (which may be deferred).
1460 */
1461 get_page(page);
1462 gnttab_end_foreign_access(ref, page);
1463 queue->grant_rx_ref[id] = INVALID_GRANT_REF;
1464
1465 kfree_skb(skb);
1466 }
1467
1468 spin_unlock_bh(&queue->rx_lock);
1469}
1470
1471static netdev_features_t xennet_fix_features(struct net_device *dev,
1472 netdev_features_t features)
1473{
1474 struct netfront_info *np = netdev_priv(dev);
1475
1476 if (features & NETIF_F_SG &&
1477 !xenbus_read_unsigned(np->xbdev->otherend, "feature-sg", 0))
1478 features &= ~NETIF_F_SG;
1479
1480 if (features & NETIF_F_IPV6_CSUM &&
1481 !xenbus_read_unsigned(np->xbdev->otherend,
1482 "feature-ipv6-csum-offload", 0))
1483 features &= ~NETIF_F_IPV6_CSUM;
1484
1485 if (features & NETIF_F_TSO &&
1486 !xenbus_read_unsigned(np->xbdev->otherend, "feature-gso-tcpv4", 0))
1487 features &= ~NETIF_F_TSO;
1488
1489 if (features & NETIF_F_TSO6 &&
1490 !xenbus_read_unsigned(np->xbdev->otherend, "feature-gso-tcpv6", 0))
1491 features &= ~NETIF_F_TSO6;
1492
1493 return features;
1494}
1495
1496static int xennet_set_features(struct net_device *dev,
1497 netdev_features_t features)
1498{
1499 if (!(features & NETIF_F_SG) && dev->mtu > ETH_DATA_LEN) {
1500 netdev_info(dev, "Reducing MTU because no SG offload");
1501 dev->mtu = ETH_DATA_LEN;
1502 }
1503
1504 return 0;
1505}
1506
1507static bool xennet_handle_tx(struct netfront_queue *queue, unsigned int *eoi)
1508{
1509 unsigned long flags;
1510
1511 if (unlikely(queue->info->broken))
1512 return false;
1513
1514 spin_lock_irqsave(&queue->tx_lock, flags);
1515 if (xennet_tx_buf_gc(queue))
1516 *eoi = 0;
1517 spin_unlock_irqrestore(&queue->tx_lock, flags);
1518
1519 return true;
1520}
1521
1522static irqreturn_t xennet_tx_interrupt(int irq, void *dev_id)
1523{
1524 unsigned int eoiflag = XEN_EOI_FLAG_SPURIOUS;
1525
1526 if (likely(xennet_handle_tx(dev_id, &eoiflag)))
1527 xen_irq_lateeoi(irq, eoiflag);
1528
1529 return IRQ_HANDLED;
1530}
1531
1532static bool xennet_handle_rx(struct netfront_queue *queue, unsigned int *eoi)
1533{
1534 unsigned int work_queued;
1535 unsigned long flags;
1536
1537 if (unlikely(queue->info->broken))
1538 return false;
1539
1540 spin_lock_irqsave(&queue->rx_cons_lock, flags);
1541 work_queued = XEN_RING_NR_UNCONSUMED_RESPONSES(&queue->rx);
1542 if (work_queued > queue->rx_rsp_unconsumed) {
1543 queue->rx_rsp_unconsumed = work_queued;
1544 *eoi = 0;
1545 } else if (unlikely(work_queued < queue->rx_rsp_unconsumed)) {
1546 const struct device *dev = &queue->info->netdev->dev;
1547
1548 spin_unlock_irqrestore(&queue->rx_cons_lock, flags);
1549 dev_alert(dev, "RX producer index going backwards\n");
1550 dev_alert(dev, "Disabled for further use\n");
1551 queue->info->broken = true;
1552 return false;
1553 }
1554 spin_unlock_irqrestore(&queue->rx_cons_lock, flags);
1555
1556 if (likely(netif_carrier_ok(queue->info->netdev) && work_queued))
1557 napi_schedule(&queue->napi);
1558
1559 return true;
1560}
1561
1562static irqreturn_t xennet_rx_interrupt(int irq, void *dev_id)
1563{
1564 unsigned int eoiflag = XEN_EOI_FLAG_SPURIOUS;
1565
1566 if (likely(xennet_handle_rx(dev_id, &eoiflag)))
1567 xen_irq_lateeoi(irq, eoiflag);
1568
1569 return IRQ_HANDLED;
1570}
1571
1572static irqreturn_t xennet_interrupt(int irq, void *dev_id)
1573{
1574 unsigned int eoiflag = XEN_EOI_FLAG_SPURIOUS;
1575
1576 if (xennet_handle_tx(dev_id, &eoiflag) &&
1577 xennet_handle_rx(dev_id, &eoiflag))
1578 xen_irq_lateeoi(irq, eoiflag);
1579
1580 return IRQ_HANDLED;
1581}
1582
1583#ifdef CONFIG_NET_POLL_CONTROLLER
1584static void xennet_poll_controller(struct net_device *dev)
1585{
1586 /* Poll each queue */
1587 struct netfront_info *info = netdev_priv(dev);
1588 unsigned int num_queues = dev->real_num_tx_queues;
1589 unsigned int i;
1590
1591 if (info->broken)
1592 return;
1593
1594 for (i = 0; i < num_queues; ++i)
1595 xennet_interrupt(0, &info->queues[i]);
1596}
1597#endif
1598
1599#define NETBACK_XDP_HEADROOM_DISABLE 0
1600#define NETBACK_XDP_HEADROOM_ENABLE 1
1601
1602static int talk_to_netback_xdp(struct netfront_info *np, int xdp)
1603{
1604 int err;
1605 unsigned short headroom;
1606
1607 headroom = xdp ? XDP_PACKET_HEADROOM : 0;
1608 err = xenbus_printf(XBT_NIL, np->xbdev->nodename,
1609 "xdp-headroom", "%hu",
1610 headroom);
1611 if (err)
1612 pr_warn("Error writing xdp-headroom\n");
1613
1614 return err;
1615}
1616
1617static int xennet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1618 struct netlink_ext_ack *extack)
1619{
1620 unsigned long max_mtu = XEN_PAGE_SIZE - XDP_PACKET_HEADROOM;
1621 struct netfront_info *np = netdev_priv(dev);
1622 struct bpf_prog *old_prog;
1623 unsigned int i, err;
1624
1625 if (dev->mtu > max_mtu) {
1626 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_mtu);
1627 return -EINVAL;
1628 }
1629
1630 if (!np->netback_has_xdp_headroom)
1631 return 0;
1632
1633 xenbus_switch_state(np->xbdev, XenbusStateReconfiguring);
1634
1635 err = talk_to_netback_xdp(np, prog ? NETBACK_XDP_HEADROOM_ENABLE :
1636 NETBACK_XDP_HEADROOM_DISABLE);
1637 if (err)
1638 return err;
1639
1640 /* avoid the race with XDP headroom adjustment */
1641 wait_event(module_wq,
1642 xenbus_read_driver_state(np->xbdev->otherend) ==
1643 XenbusStateReconfigured);
1644 np->netfront_xdp_enabled = true;
1645
1646 old_prog = rtnl_dereference(np->queues[0].xdp_prog);
1647
1648 if (prog)
1649 bpf_prog_add(prog, dev->real_num_tx_queues);
1650
1651 for (i = 0; i < dev->real_num_tx_queues; ++i)
1652 rcu_assign_pointer(np->queues[i].xdp_prog, prog);
1653
1654 if (old_prog)
1655 for (i = 0; i < dev->real_num_tx_queues; ++i)
1656 bpf_prog_put(old_prog);
1657
1658 xenbus_switch_state(np->xbdev, XenbusStateConnected);
1659
1660 return 0;
1661}
1662
1663static int xennet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
1664{
1665 struct netfront_info *np = netdev_priv(dev);
1666
1667 if (np->broken)
1668 return -ENODEV;
1669
1670 switch (xdp->command) {
1671 case XDP_SETUP_PROG:
1672 return xennet_xdp_set(dev, xdp->prog, xdp->extack);
1673 default:
1674 return -EINVAL;
1675 }
1676}
1677
1678static const struct net_device_ops xennet_netdev_ops = {
1679 .ndo_uninit = xennet_uninit,
1680 .ndo_open = xennet_open,
1681 .ndo_stop = xennet_close,
1682 .ndo_start_xmit = xennet_start_xmit,
1683 .ndo_change_mtu = xennet_change_mtu,
1684 .ndo_get_stats64 = xennet_get_stats64,
1685 .ndo_set_mac_address = eth_mac_addr,
1686 .ndo_validate_addr = eth_validate_addr,
1687 .ndo_fix_features = xennet_fix_features,
1688 .ndo_set_features = xennet_set_features,
1689 .ndo_select_queue = xennet_select_queue,
1690 .ndo_bpf = xennet_xdp,
1691 .ndo_xdp_xmit = xennet_xdp_xmit,
1692#ifdef CONFIG_NET_POLL_CONTROLLER
1693 .ndo_poll_controller = xennet_poll_controller,
1694#endif
1695};
1696
1697static void xennet_free_netdev(struct net_device *netdev)
1698{
1699 struct netfront_info *np = netdev_priv(netdev);
1700
1701 free_percpu(np->rx_stats);
1702 free_percpu(np->tx_stats);
1703 free_netdev(netdev);
1704}
1705
1706static struct net_device *xennet_create_dev(struct xenbus_device *dev)
1707{
1708 int err;
1709 struct net_device *netdev;
1710 struct netfront_info *np;
1711
1712 netdev = alloc_etherdev_mq(sizeof(struct netfront_info), xennet_max_queues);
1713 if (!netdev)
1714 return ERR_PTR(-ENOMEM);
1715
1716 np = netdev_priv(netdev);
1717 np->xbdev = dev;
1718
1719 np->queues = NULL;
1720
1721 err = -ENOMEM;
1722 np->rx_stats = netdev_alloc_pcpu_stats(struct netfront_stats);
1723 if (np->rx_stats == NULL)
1724 goto exit;
1725 np->tx_stats = netdev_alloc_pcpu_stats(struct netfront_stats);
1726 if (np->tx_stats == NULL)
1727 goto exit;
1728
1729 netdev->netdev_ops = &xennet_netdev_ops;
1730
1731 netdev->features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM |
1732 NETIF_F_GSO_ROBUST;
1733 netdev->hw_features = NETIF_F_SG |
1734 NETIF_F_IPV6_CSUM |
1735 NETIF_F_TSO | NETIF_F_TSO6;
1736
1737 /*
1738 * Assume that all hw features are available for now. This set
1739 * will be adjusted by the call to netdev_update_features() in
1740 * xennet_connect() which is the earliest point where we can
1741 * negotiate with the backend regarding supported features.
1742 */
1743 netdev->features |= netdev->hw_features;
1744 netdev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT |
1745 NETDEV_XDP_ACT_NDO_XMIT;
1746
1747 netdev->ethtool_ops = &xennet_ethtool_ops;
1748 netdev->min_mtu = ETH_MIN_MTU;
1749 netdev->max_mtu = XEN_NETIF_MAX_TX_SIZE;
1750 SET_NETDEV_DEV(netdev, &dev->dev);
1751
1752 np->netdev = netdev;
1753 np->netfront_xdp_enabled = false;
1754
1755 netif_carrier_off(netdev);
1756
1757 do {
1758 xenbus_switch_state(dev, XenbusStateInitialising);
1759 err = wait_event_timeout(module_wq,
1760 xenbus_read_driver_state(dev->otherend) !=
1761 XenbusStateClosed &&
1762 xenbus_read_driver_state(dev->otherend) !=
1763 XenbusStateUnknown, XENNET_TIMEOUT);
1764 } while (!err);
1765
1766 return netdev;
1767
1768 exit:
1769 xennet_free_netdev(netdev);
1770 return ERR_PTR(err);
1771}
1772
1773/*
1774 * Entry point to this code when a new device is created. Allocate the basic
1775 * structures and the ring buffers for communication with the backend, and
1776 * inform the backend of the appropriate details for those.
1777 */
1778static int netfront_probe(struct xenbus_device *dev,
1779 const struct xenbus_device_id *id)
1780{
1781 int err;
1782 struct net_device *netdev;
1783 struct netfront_info *info;
1784
1785 netdev = xennet_create_dev(dev);
1786 if (IS_ERR(netdev)) {
1787 err = PTR_ERR(netdev);
1788 xenbus_dev_fatal(dev, err, "creating netdev");
1789 return err;
1790 }
1791
1792 info = netdev_priv(netdev);
1793 dev_set_drvdata(&dev->dev, info);
1794#ifdef CONFIG_SYSFS
1795 info->netdev->sysfs_groups[0] = &xennet_dev_group;
1796#endif
1797
1798 return 0;
1799}
1800
1801static void xennet_end_access(int ref, void *page)
1802{
1803 /* This frees the page as a side-effect */
1804 if (ref != INVALID_GRANT_REF)
1805 gnttab_end_foreign_access(ref, virt_to_page(page));
1806}
1807
1808static void xennet_disconnect_backend(struct netfront_info *info)
1809{
1810 unsigned int i = 0;
1811 unsigned int num_queues = info->netdev->real_num_tx_queues;
1812
1813 netif_carrier_off(info->netdev);
1814
1815 for (i = 0; i < num_queues && info->queues; ++i) {
1816 struct netfront_queue *queue = &info->queues[i];
1817
1818 del_timer_sync(&queue->rx_refill_timer);
1819
1820 if (queue->tx_irq && (queue->tx_irq == queue->rx_irq))
1821 unbind_from_irqhandler(queue->tx_irq, queue);
1822 if (queue->tx_irq && (queue->tx_irq != queue->rx_irq)) {
1823 unbind_from_irqhandler(queue->tx_irq, queue);
1824 unbind_from_irqhandler(queue->rx_irq, queue);
1825 }
1826 queue->tx_evtchn = queue->rx_evtchn = 0;
1827 queue->tx_irq = queue->rx_irq = 0;
1828
1829 if (netif_running(info->netdev))
1830 napi_synchronize(&queue->napi);
1831
1832 xennet_release_tx_bufs(queue);
1833 xennet_release_rx_bufs(queue);
1834 gnttab_free_grant_references(queue->gref_tx_head);
1835 gnttab_free_grant_references(queue->gref_rx_head);
1836
1837 /* End access and free the pages */
1838 xennet_end_access(queue->tx_ring_ref, queue->tx.sring);
1839 xennet_end_access(queue->rx_ring_ref, queue->rx.sring);
1840
1841 queue->tx_ring_ref = INVALID_GRANT_REF;
1842 queue->rx_ring_ref = INVALID_GRANT_REF;
1843 queue->tx.sring = NULL;
1844 queue->rx.sring = NULL;
1845
1846 page_pool_destroy(queue->page_pool);
1847 }
1848}
1849
1850/*
1851 * We are reconnecting to the backend, due to a suspend/resume, or a backend
1852 * driver restart. We tear down our netif structure and recreate it, but
1853 * leave the device-layer structures intact so that this is transparent to the
1854 * rest of the kernel.
1855 */
1856static int netfront_resume(struct xenbus_device *dev)
1857{
1858 struct netfront_info *info = dev_get_drvdata(&dev->dev);
1859
1860 dev_dbg(&dev->dev, "%s\n", dev->nodename);
1861
1862 netif_tx_lock_bh(info->netdev);
1863 netif_device_detach(info->netdev);
1864 netif_tx_unlock_bh(info->netdev);
1865
1866 xennet_disconnect_backend(info);
1867
1868 rtnl_lock();
1869 if (info->queues)
1870 xennet_destroy_queues(info);
1871 rtnl_unlock();
1872
1873 return 0;
1874}
1875
1876static int xen_net_read_mac(struct xenbus_device *dev, u8 mac[])
1877{
1878 char *s, *e, *macstr;
1879 int i;
1880
1881 macstr = s = xenbus_read(XBT_NIL, dev->nodename, "mac", NULL);
1882 if (IS_ERR(macstr))
1883 return PTR_ERR(macstr);
1884
1885 for (i = 0; i < ETH_ALEN; i++) {
1886 mac[i] = simple_strtoul(s, &e, 16);
1887 if ((s == e) || (*e != ((i == ETH_ALEN-1) ? '\0' : ':'))) {
1888 kfree(macstr);
1889 return -ENOENT;
1890 }
1891 s = e+1;
1892 }
1893
1894 kfree(macstr);
1895 return 0;
1896}
1897
1898static int setup_netfront_single(struct netfront_queue *queue)
1899{
1900 int err;
1901
1902 err = xenbus_alloc_evtchn(queue->info->xbdev, &queue->tx_evtchn);
1903 if (err < 0)
1904 goto fail;
1905
1906 err = bind_evtchn_to_irqhandler_lateeoi(queue->tx_evtchn,
1907 xennet_interrupt, 0,
1908 queue->info->netdev->name,
1909 queue);
1910 if (err < 0)
1911 goto bind_fail;
1912 queue->rx_evtchn = queue->tx_evtchn;
1913 queue->rx_irq = queue->tx_irq = err;
1914
1915 return 0;
1916
1917bind_fail:
1918 xenbus_free_evtchn(queue->info->xbdev, queue->tx_evtchn);
1919 queue->tx_evtchn = 0;
1920fail:
1921 return err;
1922}
1923
1924static int setup_netfront_split(struct netfront_queue *queue)
1925{
1926 int err;
1927
1928 err = xenbus_alloc_evtchn(queue->info->xbdev, &queue->tx_evtchn);
1929 if (err < 0)
1930 goto fail;
1931 err = xenbus_alloc_evtchn(queue->info->xbdev, &queue->rx_evtchn);
1932 if (err < 0)
1933 goto alloc_rx_evtchn_fail;
1934
1935 snprintf(queue->tx_irq_name, sizeof(queue->tx_irq_name),
1936 "%s-tx", queue->name);
1937 err = bind_evtchn_to_irqhandler_lateeoi(queue->tx_evtchn,
1938 xennet_tx_interrupt, 0,
1939 queue->tx_irq_name, queue);
1940 if (err < 0)
1941 goto bind_tx_fail;
1942 queue->tx_irq = err;
1943
1944 snprintf(queue->rx_irq_name, sizeof(queue->rx_irq_name),
1945 "%s-rx", queue->name);
1946 err = bind_evtchn_to_irqhandler_lateeoi(queue->rx_evtchn,
1947 xennet_rx_interrupt, 0,
1948 queue->rx_irq_name, queue);
1949 if (err < 0)
1950 goto bind_rx_fail;
1951 queue->rx_irq = err;
1952
1953 return 0;
1954
1955bind_rx_fail:
1956 unbind_from_irqhandler(queue->tx_irq, queue);
1957 queue->tx_irq = 0;
1958bind_tx_fail:
1959 xenbus_free_evtchn(queue->info->xbdev, queue->rx_evtchn);
1960 queue->rx_evtchn = 0;
1961alloc_rx_evtchn_fail:
1962 xenbus_free_evtchn(queue->info->xbdev, queue->tx_evtchn);
1963 queue->tx_evtchn = 0;
1964fail:
1965 return err;
1966}
1967
1968static int setup_netfront(struct xenbus_device *dev,
1969 struct netfront_queue *queue, unsigned int feature_split_evtchn)
1970{
1971 struct xen_netif_tx_sring *txs;
1972 struct xen_netif_rx_sring *rxs;
1973 int err;
1974
1975 queue->tx_ring_ref = INVALID_GRANT_REF;
1976 queue->rx_ring_ref = INVALID_GRANT_REF;
1977 queue->rx.sring = NULL;
1978 queue->tx.sring = NULL;
1979
1980 err = xenbus_setup_ring(dev, GFP_NOIO | __GFP_HIGH, (void **)&txs,
1981 1, &queue->tx_ring_ref);
1982 if (err)
1983 goto fail;
1984
1985 XEN_FRONT_RING_INIT(&queue->tx, txs, XEN_PAGE_SIZE);
1986
1987 err = xenbus_setup_ring(dev, GFP_NOIO | __GFP_HIGH, (void **)&rxs,
1988 1, &queue->rx_ring_ref);
1989 if (err)
1990 goto fail;
1991
1992 XEN_FRONT_RING_INIT(&queue->rx, rxs, XEN_PAGE_SIZE);
1993
1994 if (feature_split_evtchn)
1995 err = setup_netfront_split(queue);
1996 /* setup single event channel if
1997 * a) feature-split-event-channels == 0
1998 * b) feature-split-event-channels == 1 but failed to setup
1999 */
2000 if (!feature_split_evtchn || err)
2001 err = setup_netfront_single(queue);
2002
2003 if (err)
2004 goto fail;
2005
2006 return 0;
2007
2008 fail:
2009 xenbus_teardown_ring((void **)&queue->rx.sring, 1, &queue->rx_ring_ref);
2010 xenbus_teardown_ring((void **)&queue->tx.sring, 1, &queue->tx_ring_ref);
2011
2012 return err;
2013}
2014
2015/* Queue-specific initialisation
2016 * This used to be done in xennet_create_dev() but must now
2017 * be run per-queue.
2018 */
2019static int xennet_init_queue(struct netfront_queue *queue)
2020{
2021 unsigned short i;
2022 int err = 0;
2023 char *devid;
2024
2025 spin_lock_init(&queue->tx_lock);
2026 spin_lock_init(&queue->rx_lock);
2027 spin_lock_init(&queue->rx_cons_lock);
2028
2029 timer_setup(&queue->rx_refill_timer, rx_refill_timeout, 0);
2030
2031 devid = strrchr(queue->info->xbdev->nodename, '/') + 1;
2032 snprintf(queue->name, sizeof(queue->name), "vif%s-q%u",
2033 devid, queue->id);
2034
2035 /* Initialise tx_skb_freelist as a free chain containing every entry. */
2036 queue->tx_skb_freelist = 0;
2037 queue->tx_pend_queue = TX_LINK_NONE;
2038 for (i = 0; i < NET_TX_RING_SIZE; i++) {
2039 queue->tx_link[i] = i + 1;
2040 queue->grant_tx_ref[i] = INVALID_GRANT_REF;
2041 queue->grant_tx_page[i] = NULL;
2042 }
2043 queue->tx_link[NET_TX_RING_SIZE - 1] = TX_LINK_NONE;
2044
2045 /* Clear out rx_skbs */
2046 for (i = 0; i < NET_RX_RING_SIZE; i++) {
2047 queue->rx_skbs[i] = NULL;
2048 queue->grant_rx_ref[i] = INVALID_GRANT_REF;
2049 }
2050
2051 /* A grant for every tx ring slot */
2052 if (gnttab_alloc_grant_references(NET_TX_RING_SIZE,
2053 &queue->gref_tx_head) < 0) {
2054 pr_alert("can't alloc tx grant refs\n");
2055 err = -ENOMEM;
2056 goto exit;
2057 }
2058
2059 /* A grant for every rx ring slot */
2060 if (gnttab_alloc_grant_references(NET_RX_RING_SIZE,
2061 &queue->gref_rx_head) < 0) {
2062 pr_alert("can't alloc rx grant refs\n");
2063 err = -ENOMEM;
2064 goto exit_free_tx;
2065 }
2066
2067 return 0;
2068
2069 exit_free_tx:
2070 gnttab_free_grant_references(queue->gref_tx_head);
2071 exit:
2072 return err;
2073}
2074
2075static int write_queue_xenstore_keys(struct netfront_queue *queue,
2076 struct xenbus_transaction *xbt, int write_hierarchical)
2077{
2078 /* Write the queue-specific keys into XenStore in the traditional
2079 * way for a single queue, or in a queue subkeys for multiple
2080 * queues.
2081 */
2082 struct xenbus_device *dev = queue->info->xbdev;
2083 int err;
2084 const char *message;
2085 char *path;
2086 size_t pathsize;
2087
2088 /* Choose the correct place to write the keys */
2089 if (write_hierarchical) {
2090 pathsize = strlen(dev->nodename) + 10;
2091 path = kzalloc(pathsize, GFP_KERNEL);
2092 if (!path) {
2093 err = -ENOMEM;
2094 message = "out of memory while writing ring references";
2095 goto error;
2096 }
2097 snprintf(path, pathsize, "%s/queue-%u",
2098 dev->nodename, queue->id);
2099 } else {
2100 path = (char *)dev->nodename;
2101 }
2102
2103 /* Write ring references */
2104 err = xenbus_printf(*xbt, path, "tx-ring-ref", "%u",
2105 queue->tx_ring_ref);
2106 if (err) {
2107 message = "writing tx-ring-ref";
2108 goto error;
2109 }
2110
2111 err = xenbus_printf(*xbt, path, "rx-ring-ref", "%u",
2112 queue->rx_ring_ref);
2113 if (err) {
2114 message = "writing rx-ring-ref";
2115 goto error;
2116 }
2117
2118 /* Write event channels; taking into account both shared
2119 * and split event channel scenarios.
2120 */
2121 if (queue->tx_evtchn == queue->rx_evtchn) {
2122 /* Shared event channel */
2123 err = xenbus_printf(*xbt, path,
2124 "event-channel", "%u", queue->tx_evtchn);
2125 if (err) {
2126 message = "writing event-channel";
2127 goto error;
2128 }
2129 } else {
2130 /* Split event channels */
2131 err = xenbus_printf(*xbt, path,
2132 "event-channel-tx", "%u", queue->tx_evtchn);
2133 if (err) {
2134 message = "writing event-channel-tx";
2135 goto error;
2136 }
2137
2138 err = xenbus_printf(*xbt, path,
2139 "event-channel-rx", "%u", queue->rx_evtchn);
2140 if (err) {
2141 message = "writing event-channel-rx";
2142 goto error;
2143 }
2144 }
2145
2146 if (write_hierarchical)
2147 kfree(path);
2148 return 0;
2149
2150error:
2151 if (write_hierarchical)
2152 kfree(path);
2153 xenbus_dev_fatal(dev, err, "%s", message);
2154 return err;
2155}
2156
2157
2158
2159static int xennet_create_page_pool(struct netfront_queue *queue)
2160{
2161 int err;
2162 struct page_pool_params pp_params = {
2163 .order = 0,
2164 .flags = 0,
2165 .pool_size = NET_RX_RING_SIZE,
2166 .nid = NUMA_NO_NODE,
2167 .dev = &queue->info->netdev->dev,
2168 .offset = XDP_PACKET_HEADROOM,
2169 .max_len = XEN_PAGE_SIZE - XDP_PACKET_HEADROOM,
2170 };
2171
2172 queue->page_pool = page_pool_create(&pp_params);
2173 if (IS_ERR(queue->page_pool)) {
2174 err = PTR_ERR(queue->page_pool);
2175 queue->page_pool = NULL;
2176 return err;
2177 }
2178
2179 err = xdp_rxq_info_reg(&queue->xdp_rxq, queue->info->netdev,
2180 queue->id, 0);
2181 if (err) {
2182 netdev_err(queue->info->netdev, "xdp_rxq_info_reg failed\n");
2183 goto err_free_pp;
2184 }
2185
2186 err = xdp_rxq_info_reg_mem_model(&queue->xdp_rxq,
2187 MEM_TYPE_PAGE_POOL, queue->page_pool);
2188 if (err) {
2189 netdev_err(queue->info->netdev, "xdp_rxq_info_reg_mem_model failed\n");
2190 goto err_unregister_rxq;
2191 }
2192 return 0;
2193
2194err_unregister_rxq:
2195 xdp_rxq_info_unreg(&queue->xdp_rxq);
2196err_free_pp:
2197 page_pool_destroy(queue->page_pool);
2198 queue->page_pool = NULL;
2199 return err;
2200}
2201
2202static int xennet_create_queues(struct netfront_info *info,
2203 unsigned int *num_queues)
2204{
2205 unsigned int i;
2206 int ret;
2207
2208 info->queues = kcalloc(*num_queues, sizeof(struct netfront_queue),
2209 GFP_KERNEL);
2210 if (!info->queues)
2211 return -ENOMEM;
2212
2213 for (i = 0; i < *num_queues; i++) {
2214 struct netfront_queue *queue = &info->queues[i];
2215
2216 queue->id = i;
2217 queue->info = info;
2218
2219 ret = xennet_init_queue(queue);
2220 if (ret < 0) {
2221 dev_warn(&info->xbdev->dev,
2222 "only created %d queues\n", i);
2223 *num_queues = i;
2224 break;
2225 }
2226
2227 /* use page pool recycling instead of buddy allocator */
2228 ret = xennet_create_page_pool(queue);
2229 if (ret < 0) {
2230 dev_err(&info->xbdev->dev, "can't allocate page pool\n");
2231 *num_queues = i;
2232 return ret;
2233 }
2234
2235 netif_napi_add(queue->info->netdev, &queue->napi, xennet_poll);
2236 if (netif_running(info->netdev))
2237 napi_enable(&queue->napi);
2238 }
2239
2240 netif_set_real_num_tx_queues(info->netdev, *num_queues);
2241
2242 if (*num_queues == 0) {
2243 dev_err(&info->xbdev->dev, "no queues\n");
2244 return -EINVAL;
2245 }
2246 return 0;
2247}
2248
2249/* Common code used when first setting up, and when resuming. */
2250static int talk_to_netback(struct xenbus_device *dev,
2251 struct netfront_info *info)
2252{
2253 const char *message;
2254 struct xenbus_transaction xbt;
2255 int err;
2256 unsigned int feature_split_evtchn;
2257 unsigned int i = 0;
2258 unsigned int max_queues = 0;
2259 struct netfront_queue *queue = NULL;
2260 unsigned int num_queues = 1;
2261 u8 addr[ETH_ALEN];
2262
2263 info->netdev->irq = 0;
2264
2265 /* Check if backend is trusted. */
2266 info->bounce = !xennet_trusted ||
2267 !xenbus_read_unsigned(dev->nodename, "trusted", 1);
2268
2269 /* Check if backend supports multiple queues */
2270 max_queues = xenbus_read_unsigned(info->xbdev->otherend,
2271 "multi-queue-max-queues", 1);
2272 num_queues = min(max_queues, xennet_max_queues);
2273
2274 /* Check feature-split-event-channels */
2275 feature_split_evtchn = xenbus_read_unsigned(info->xbdev->otherend,
2276 "feature-split-event-channels", 0);
2277
2278 /* Read mac addr. */
2279 err = xen_net_read_mac(dev, addr);
2280 if (err) {
2281 xenbus_dev_fatal(dev, err, "parsing %s/mac", dev->nodename);
2282 goto out_unlocked;
2283 }
2284 eth_hw_addr_set(info->netdev, addr);
2285
2286 info->netback_has_xdp_headroom = xenbus_read_unsigned(info->xbdev->otherend,
2287 "feature-xdp-headroom", 0);
2288 if (info->netback_has_xdp_headroom) {
2289 /* set the current xen-netfront xdp state */
2290 err = talk_to_netback_xdp(info, info->netfront_xdp_enabled ?
2291 NETBACK_XDP_HEADROOM_ENABLE :
2292 NETBACK_XDP_HEADROOM_DISABLE);
2293 if (err)
2294 goto out_unlocked;
2295 }
2296
2297 rtnl_lock();
2298 if (info->queues)
2299 xennet_destroy_queues(info);
2300
2301 /* For the case of a reconnect reset the "broken" indicator. */
2302 info->broken = false;
2303
2304 err = xennet_create_queues(info, &num_queues);
2305 if (err < 0) {
2306 xenbus_dev_fatal(dev, err, "creating queues");
2307 kfree(info->queues);
2308 info->queues = NULL;
2309 goto out;
2310 }
2311 rtnl_unlock();
2312
2313 /* Create shared ring, alloc event channel -- for each queue */
2314 for (i = 0; i < num_queues; ++i) {
2315 queue = &info->queues[i];
2316 err = setup_netfront(dev, queue, feature_split_evtchn);
2317 if (err)
2318 goto destroy_ring;
2319 }
2320
2321again:
2322 err = xenbus_transaction_start(&xbt);
2323 if (err) {
2324 xenbus_dev_fatal(dev, err, "starting transaction");
2325 goto destroy_ring;
2326 }
2327
2328 if (xenbus_exists(XBT_NIL,
2329 info->xbdev->otherend, "multi-queue-max-queues")) {
2330 /* Write the number of queues */
2331 err = xenbus_printf(xbt, dev->nodename,
2332 "multi-queue-num-queues", "%u", num_queues);
2333 if (err) {
2334 message = "writing multi-queue-num-queues";
2335 goto abort_transaction_no_dev_fatal;
2336 }
2337 }
2338
2339 if (num_queues == 1) {
2340 err = write_queue_xenstore_keys(&info->queues[0], &xbt, 0); /* flat */
2341 if (err)
2342 goto abort_transaction_no_dev_fatal;
2343 } else {
2344 /* Write the keys for each queue */
2345 for (i = 0; i < num_queues; ++i) {
2346 queue = &info->queues[i];
2347 err = write_queue_xenstore_keys(queue, &xbt, 1); /* hierarchical */
2348 if (err)
2349 goto abort_transaction_no_dev_fatal;
2350 }
2351 }
2352
2353 /* The remaining keys are not queue-specific */
2354 err = xenbus_printf(xbt, dev->nodename, "request-rx-copy", "%u",
2355 1);
2356 if (err) {
2357 message = "writing request-rx-copy";
2358 goto abort_transaction;
2359 }
2360
2361 err = xenbus_printf(xbt, dev->nodename, "feature-rx-notify", "%d", 1);
2362 if (err) {
2363 message = "writing feature-rx-notify";
2364 goto abort_transaction;
2365 }
2366
2367 err = xenbus_printf(xbt, dev->nodename, "feature-sg", "%d", 1);
2368 if (err) {
2369 message = "writing feature-sg";
2370 goto abort_transaction;
2371 }
2372
2373 err = xenbus_printf(xbt, dev->nodename, "feature-gso-tcpv4", "%d", 1);
2374 if (err) {
2375 message = "writing feature-gso-tcpv4";
2376 goto abort_transaction;
2377 }
2378
2379 err = xenbus_write(xbt, dev->nodename, "feature-gso-tcpv6", "1");
2380 if (err) {
2381 message = "writing feature-gso-tcpv6";
2382 goto abort_transaction;
2383 }
2384
2385 err = xenbus_write(xbt, dev->nodename, "feature-ipv6-csum-offload",
2386 "1");
2387 if (err) {
2388 message = "writing feature-ipv6-csum-offload";
2389 goto abort_transaction;
2390 }
2391
2392 err = xenbus_transaction_end(xbt, 0);
2393 if (err) {
2394 if (err == -EAGAIN)
2395 goto again;
2396 xenbus_dev_fatal(dev, err, "completing transaction");
2397 goto destroy_ring;
2398 }
2399
2400 return 0;
2401
2402 abort_transaction:
2403 xenbus_dev_fatal(dev, err, "%s", message);
2404abort_transaction_no_dev_fatal:
2405 xenbus_transaction_end(xbt, 1);
2406 destroy_ring:
2407 xennet_disconnect_backend(info);
2408 rtnl_lock();
2409 xennet_destroy_queues(info);
2410 out:
2411 rtnl_unlock();
2412out_unlocked:
2413 device_unregister(&dev->dev);
2414 return err;
2415}
2416
2417static int xennet_connect(struct net_device *dev)
2418{
2419 struct netfront_info *np = netdev_priv(dev);
2420 unsigned int num_queues = 0;
2421 int err;
2422 unsigned int j = 0;
2423 struct netfront_queue *queue = NULL;
2424
2425 if (!xenbus_read_unsigned(np->xbdev->otherend, "feature-rx-copy", 0)) {
2426 dev_info(&dev->dev,
2427 "backend does not support copying receive path\n");
2428 return -ENODEV;
2429 }
2430
2431 err = talk_to_netback(np->xbdev, np);
2432 if (err)
2433 return err;
2434 if (np->netback_has_xdp_headroom)
2435 pr_info("backend supports XDP headroom\n");
2436 if (np->bounce)
2437 dev_info(&np->xbdev->dev,
2438 "bouncing transmitted data to zeroed pages\n");
2439
2440 /* talk_to_netback() sets the correct number of queues */
2441 num_queues = dev->real_num_tx_queues;
2442
2443 if (dev->reg_state == NETREG_UNINITIALIZED) {
2444 err = register_netdev(dev);
2445 if (err) {
2446 pr_warn("%s: register_netdev err=%d\n", __func__, err);
2447 device_unregister(&np->xbdev->dev);
2448 return err;
2449 }
2450 }
2451
2452 rtnl_lock();
2453 netdev_update_features(dev);
2454 rtnl_unlock();
2455
2456 /*
2457 * All public and private state should now be sane. Get
2458 * ready to start sending and receiving packets and give the driver
2459 * domain a kick because we've probably just requeued some
2460 * packets.
2461 */
2462 netif_tx_lock_bh(np->netdev);
2463 netif_device_attach(np->netdev);
2464 netif_tx_unlock_bh(np->netdev);
2465
2466 netif_carrier_on(np->netdev);
2467 for (j = 0; j < num_queues; ++j) {
2468 queue = &np->queues[j];
2469
2470 notify_remote_via_irq(queue->tx_irq);
2471 if (queue->tx_irq != queue->rx_irq)
2472 notify_remote_via_irq(queue->rx_irq);
2473
2474 spin_lock_bh(&queue->rx_lock);
2475 xennet_alloc_rx_buffers(queue);
2476 spin_unlock_bh(&queue->rx_lock);
2477 }
2478
2479 return 0;
2480}
2481
2482/*
2483 * Callback received when the backend's state changes.
2484 */
2485static void netback_changed(struct xenbus_device *dev,
2486 enum xenbus_state backend_state)
2487{
2488 struct netfront_info *np = dev_get_drvdata(&dev->dev);
2489 struct net_device *netdev = np->netdev;
2490
2491 dev_dbg(&dev->dev, "%s\n", xenbus_strstate(backend_state));
2492
2493 wake_up_all(&module_wq);
2494
2495 switch (backend_state) {
2496 case XenbusStateInitialising:
2497 case XenbusStateInitialised:
2498 case XenbusStateReconfiguring:
2499 case XenbusStateReconfigured:
2500 case XenbusStateUnknown:
2501 break;
2502
2503 case XenbusStateInitWait:
2504 if (dev->state != XenbusStateInitialising)
2505 break;
2506 if (xennet_connect(netdev) != 0)
2507 break;
2508 xenbus_switch_state(dev, XenbusStateConnected);
2509 break;
2510
2511 case XenbusStateConnected:
2512 netdev_notify_peers(netdev);
2513 break;
2514
2515 case XenbusStateClosed:
2516 if (dev->state == XenbusStateClosed)
2517 break;
2518 fallthrough; /* Missed the backend's CLOSING state */
2519 case XenbusStateClosing:
2520 xenbus_frontend_closed(dev);
2521 break;
2522 }
2523}
2524
2525static const struct xennet_stat {
2526 char name[ETH_GSTRING_LEN];
2527 u16 offset;
2528} xennet_stats[] = {
2529 {
2530 "rx_gso_checksum_fixup",
2531 offsetof(struct netfront_info, rx_gso_checksum_fixup)
2532 },
2533};
2534
2535static int xennet_get_sset_count(struct net_device *dev, int string_set)
2536{
2537 switch (string_set) {
2538 case ETH_SS_STATS:
2539 return ARRAY_SIZE(xennet_stats);
2540 default:
2541 return -EINVAL;
2542 }
2543}
2544
2545static void xennet_get_ethtool_stats(struct net_device *dev,
2546 struct ethtool_stats *stats, u64 * data)
2547{
2548 void *np = netdev_priv(dev);
2549 int i;
2550
2551 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++)
2552 data[i] = atomic_read((atomic_t *)(np + xennet_stats[i].offset));
2553}
2554
2555static void xennet_get_strings(struct net_device *dev, u32 stringset, u8 * data)
2556{
2557 int i;
2558
2559 switch (stringset) {
2560 case ETH_SS_STATS:
2561 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++)
2562 memcpy(data + i * ETH_GSTRING_LEN,
2563 xennet_stats[i].name, ETH_GSTRING_LEN);
2564 break;
2565 }
2566}
2567
2568static const struct ethtool_ops xennet_ethtool_ops =
2569{
2570 .get_link = ethtool_op_get_link,
2571
2572 .get_sset_count = xennet_get_sset_count,
2573 .get_ethtool_stats = xennet_get_ethtool_stats,
2574 .get_strings = xennet_get_strings,
2575 .get_ts_info = ethtool_op_get_ts_info,
2576};
2577
2578#ifdef CONFIG_SYSFS
2579static ssize_t show_rxbuf(struct device *dev,
2580 struct device_attribute *attr, char *buf)
2581{
2582 return sprintf(buf, "%lu\n", NET_RX_RING_SIZE);
2583}
2584
2585static ssize_t store_rxbuf(struct device *dev,
2586 struct device_attribute *attr,
2587 const char *buf, size_t len)
2588{
2589 char *endp;
2590
2591 if (!capable(CAP_NET_ADMIN))
2592 return -EPERM;
2593
2594 simple_strtoul(buf, &endp, 0);
2595 if (endp == buf)
2596 return -EBADMSG;
2597
2598 /* rxbuf_min and rxbuf_max are no longer configurable. */
2599
2600 return len;
2601}
2602
2603static DEVICE_ATTR(rxbuf_min, 0644, show_rxbuf, store_rxbuf);
2604static DEVICE_ATTR(rxbuf_max, 0644, show_rxbuf, store_rxbuf);
2605static DEVICE_ATTR(rxbuf_cur, 0444, show_rxbuf, NULL);
2606
2607static struct attribute *xennet_dev_attrs[] = {
2608 &dev_attr_rxbuf_min.attr,
2609 &dev_attr_rxbuf_max.attr,
2610 &dev_attr_rxbuf_cur.attr,
2611 NULL
2612};
2613
2614static const struct attribute_group xennet_dev_group = {
2615 .attrs = xennet_dev_attrs
2616};
2617#endif /* CONFIG_SYSFS */
2618
2619static void xennet_bus_close(struct xenbus_device *dev)
2620{
2621 int ret;
2622
2623 if (xenbus_read_driver_state(dev->otherend) == XenbusStateClosed)
2624 return;
2625 do {
2626 xenbus_switch_state(dev, XenbusStateClosing);
2627 ret = wait_event_timeout(module_wq,
2628 xenbus_read_driver_state(dev->otherend) ==
2629 XenbusStateClosing ||
2630 xenbus_read_driver_state(dev->otherend) ==
2631 XenbusStateClosed ||
2632 xenbus_read_driver_state(dev->otherend) ==
2633 XenbusStateUnknown,
2634 XENNET_TIMEOUT);
2635 } while (!ret);
2636
2637 if (xenbus_read_driver_state(dev->otherend) == XenbusStateClosed)
2638 return;
2639
2640 do {
2641 xenbus_switch_state(dev, XenbusStateClosed);
2642 ret = wait_event_timeout(module_wq,
2643 xenbus_read_driver_state(dev->otherend) ==
2644 XenbusStateClosed ||
2645 xenbus_read_driver_state(dev->otherend) ==
2646 XenbusStateUnknown,
2647 XENNET_TIMEOUT);
2648 } while (!ret);
2649}
2650
2651static void xennet_remove(struct xenbus_device *dev)
2652{
2653 struct netfront_info *info = dev_get_drvdata(&dev->dev);
2654
2655 xennet_bus_close(dev);
2656 xennet_disconnect_backend(info);
2657
2658 if (info->netdev->reg_state == NETREG_REGISTERED)
2659 unregister_netdev(info->netdev);
2660
2661 if (info->queues) {
2662 rtnl_lock();
2663 xennet_destroy_queues(info);
2664 rtnl_unlock();
2665 }
2666 xennet_free_netdev(info->netdev);
2667}
2668
2669static const struct xenbus_device_id netfront_ids[] = {
2670 { "vif" },
2671 { "" }
2672};
2673
2674static struct xenbus_driver netfront_driver = {
2675 .ids = netfront_ids,
2676 .probe = netfront_probe,
2677 .remove = xennet_remove,
2678 .resume = netfront_resume,
2679 .otherend_changed = netback_changed,
2680};
2681
2682static int __init netif_init(void)
2683{
2684 if (!xen_domain())
2685 return -ENODEV;
2686
2687 if (!xen_has_pv_nic_devices())
2688 return -ENODEV;
2689
2690 pr_info("Initialising Xen virtual ethernet driver\n");
2691
2692 /* Allow as many queues as there are CPUs inut max. 8 if user has not
2693 * specified a value.
2694 */
2695 if (xennet_max_queues == 0)
2696 xennet_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
2697 num_online_cpus());
2698
2699 return xenbus_register_frontend(&netfront_driver);
2700}
2701module_init(netif_init);
2702
2703
2704static void __exit netif_exit(void)
2705{
2706 xenbus_unregister_driver(&netfront_driver);
2707}
2708module_exit(netif_exit);
2709
2710MODULE_DESCRIPTION("Xen virtual network device frontend");
2711MODULE_LICENSE("GPL");
2712MODULE_ALIAS("xen:vif");
2713MODULE_ALIAS("xennet");
1/*
2 * Virtual network driver for conversing with remote driver backends.
3 *
4 * Copyright (c) 2002-2005, K A Fraser
5 * Copyright (c) 2005, XenSource Ltd
6 *
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License version 2
9 * as published by the Free Software Foundation; or, when distributed
10 * separately from the Linux kernel or incorporated into other
11 * software packages, subject to the following license:
12 *
13 * Permission is hereby granted, free of charge, to any person obtaining a copy
14 * of this source file (the "Software"), to deal in the Software without
15 * restriction, including without limitation the rights to use, copy, modify,
16 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
17 * and to permit persons to whom the Software is furnished to do so, subject to
18 * the following conditions:
19 *
20 * The above copyright notice and this permission notice shall be included in
21 * all copies or substantial portions of the Software.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
24 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
25 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
26 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
27 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
28 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
29 * IN THE SOFTWARE.
30 */
31
32#include <linux/module.h>
33#include <linux/kernel.h>
34#include <linux/netdevice.h>
35#include <linux/etherdevice.h>
36#include <linux/skbuff.h>
37#include <linux/ethtool.h>
38#include <linux/if_ether.h>
39#include <linux/tcp.h>
40#include <linux/udp.h>
41#include <linux/moduleparam.h>
42#include <linux/mm.h>
43#include <linux/slab.h>
44#include <net/ip.h>
45
46#include <xen/xen.h>
47#include <xen/xenbus.h>
48#include <xen/events.h>
49#include <xen/page.h>
50#include <xen/platform_pci.h>
51#include <xen/grant_table.h>
52
53#include <xen/interface/io/netif.h>
54#include <xen/interface/memory.h>
55#include <xen/interface/grant_table.h>
56
57static const struct ethtool_ops xennet_ethtool_ops;
58
59struct netfront_cb {
60 struct page *page;
61 unsigned offset;
62};
63
64#define NETFRONT_SKB_CB(skb) ((struct netfront_cb *)((skb)->cb))
65
66#define RX_COPY_THRESHOLD 256
67
68#define GRANT_INVALID_REF 0
69
70#define NET_TX_RING_SIZE __CONST_RING_SIZE(xen_netif_tx, PAGE_SIZE)
71#define NET_RX_RING_SIZE __CONST_RING_SIZE(xen_netif_rx, PAGE_SIZE)
72#define TX_MAX_TARGET min_t(int, NET_TX_RING_SIZE, 256)
73
74struct netfront_stats {
75 u64 rx_packets;
76 u64 tx_packets;
77 u64 rx_bytes;
78 u64 tx_bytes;
79 struct u64_stats_sync syncp;
80};
81
82struct netfront_info {
83 struct list_head list;
84 struct net_device *netdev;
85
86 struct napi_struct napi;
87
88 unsigned int evtchn;
89 struct xenbus_device *xbdev;
90
91 spinlock_t tx_lock;
92 struct xen_netif_tx_front_ring tx;
93 int tx_ring_ref;
94
95 /*
96 * {tx,rx}_skbs store outstanding skbuffs. Free tx_skb entries
97 * are linked from tx_skb_freelist through skb_entry.link.
98 *
99 * NB. Freelist index entries are always going to be less than
100 * PAGE_OFFSET, whereas pointers to skbs will always be equal or
101 * greater than PAGE_OFFSET: we use this property to distinguish
102 * them.
103 */
104 union skb_entry {
105 struct sk_buff *skb;
106 unsigned long link;
107 } tx_skbs[NET_TX_RING_SIZE];
108 grant_ref_t gref_tx_head;
109 grant_ref_t grant_tx_ref[NET_TX_RING_SIZE];
110 unsigned tx_skb_freelist;
111
112 spinlock_t rx_lock ____cacheline_aligned_in_smp;
113 struct xen_netif_rx_front_ring rx;
114 int rx_ring_ref;
115
116 /* Receive-ring batched refills. */
117#define RX_MIN_TARGET 8
118#define RX_DFL_MIN_TARGET 64
119#define RX_MAX_TARGET min_t(int, NET_RX_RING_SIZE, 256)
120 unsigned rx_min_target, rx_max_target, rx_target;
121 struct sk_buff_head rx_batch;
122
123 struct timer_list rx_refill_timer;
124
125 struct sk_buff *rx_skbs[NET_RX_RING_SIZE];
126 grant_ref_t gref_rx_head;
127 grant_ref_t grant_rx_ref[NET_RX_RING_SIZE];
128
129 unsigned long rx_pfn_array[NET_RX_RING_SIZE];
130 struct multicall_entry rx_mcl[NET_RX_RING_SIZE+1];
131 struct mmu_update rx_mmu[NET_RX_RING_SIZE];
132
133 /* Statistics */
134 struct netfront_stats __percpu *stats;
135
136 unsigned long rx_gso_checksum_fixup;
137};
138
139struct netfront_rx_info {
140 struct xen_netif_rx_response rx;
141 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
142};
143
144static void skb_entry_set_link(union skb_entry *list, unsigned short id)
145{
146 list->link = id;
147}
148
149static int skb_entry_is_link(const union skb_entry *list)
150{
151 BUILD_BUG_ON(sizeof(list->skb) != sizeof(list->link));
152 return (unsigned long)list->skb < PAGE_OFFSET;
153}
154
155/*
156 * Access macros for acquiring freeing slots in tx_skbs[].
157 */
158
159static void add_id_to_freelist(unsigned *head, union skb_entry *list,
160 unsigned short id)
161{
162 skb_entry_set_link(&list[id], *head);
163 *head = id;
164}
165
166static unsigned short get_id_from_freelist(unsigned *head,
167 union skb_entry *list)
168{
169 unsigned int id = *head;
170 *head = list[id].link;
171 return id;
172}
173
174static int xennet_rxidx(RING_IDX idx)
175{
176 return idx & (NET_RX_RING_SIZE - 1);
177}
178
179static struct sk_buff *xennet_get_rx_skb(struct netfront_info *np,
180 RING_IDX ri)
181{
182 int i = xennet_rxidx(ri);
183 struct sk_buff *skb = np->rx_skbs[i];
184 np->rx_skbs[i] = NULL;
185 return skb;
186}
187
188static grant_ref_t xennet_get_rx_ref(struct netfront_info *np,
189 RING_IDX ri)
190{
191 int i = xennet_rxidx(ri);
192 grant_ref_t ref = np->grant_rx_ref[i];
193 np->grant_rx_ref[i] = GRANT_INVALID_REF;
194 return ref;
195}
196
197#ifdef CONFIG_SYSFS
198static int xennet_sysfs_addif(struct net_device *netdev);
199static void xennet_sysfs_delif(struct net_device *netdev);
200#else /* !CONFIG_SYSFS */
201#define xennet_sysfs_addif(dev) (0)
202#define xennet_sysfs_delif(dev) do { } while (0)
203#endif
204
205static bool xennet_can_sg(struct net_device *dev)
206{
207 return dev->features & NETIF_F_SG;
208}
209
210
211static void rx_refill_timeout(unsigned long data)
212{
213 struct net_device *dev = (struct net_device *)data;
214 struct netfront_info *np = netdev_priv(dev);
215 napi_schedule(&np->napi);
216}
217
218static int netfront_tx_slot_available(struct netfront_info *np)
219{
220 return (np->tx.req_prod_pvt - np->tx.rsp_cons) <
221 (TX_MAX_TARGET - MAX_SKB_FRAGS - 2);
222}
223
224static void xennet_maybe_wake_tx(struct net_device *dev)
225{
226 struct netfront_info *np = netdev_priv(dev);
227
228 if (unlikely(netif_queue_stopped(dev)) &&
229 netfront_tx_slot_available(np) &&
230 likely(netif_running(dev)))
231 netif_wake_queue(dev);
232}
233
234static void xennet_alloc_rx_buffers(struct net_device *dev)
235{
236 unsigned short id;
237 struct netfront_info *np = netdev_priv(dev);
238 struct sk_buff *skb;
239 struct page *page;
240 int i, batch_target, notify;
241 RING_IDX req_prod = np->rx.req_prod_pvt;
242 grant_ref_t ref;
243 unsigned long pfn;
244 void *vaddr;
245 struct xen_netif_rx_request *req;
246
247 if (unlikely(!netif_carrier_ok(dev)))
248 return;
249
250 /*
251 * Allocate skbuffs greedily, even though we batch updates to the
252 * receive ring. This creates a less bursty demand on the memory
253 * allocator, so should reduce the chance of failed allocation requests
254 * both for ourself and for other kernel subsystems.
255 */
256 batch_target = np->rx_target - (req_prod - np->rx.rsp_cons);
257 for (i = skb_queue_len(&np->rx_batch); i < batch_target; i++) {
258 skb = __netdev_alloc_skb(dev, RX_COPY_THRESHOLD + NET_IP_ALIGN,
259 GFP_ATOMIC | __GFP_NOWARN);
260 if (unlikely(!skb))
261 goto no_skb;
262
263 /* Align ip header to a 16 bytes boundary */
264 skb_reserve(skb, NET_IP_ALIGN);
265
266 page = alloc_page(GFP_ATOMIC | __GFP_NOWARN);
267 if (!page) {
268 kfree_skb(skb);
269no_skb:
270 /* Any skbuffs queued for refill? Force them out. */
271 if (i != 0)
272 goto refill;
273 /* Could not allocate any skbuffs. Try again later. */
274 mod_timer(&np->rx_refill_timer,
275 jiffies + (HZ/10));
276 break;
277 }
278
279 __skb_fill_page_desc(skb, 0, page, 0, 0);
280 skb_shinfo(skb)->nr_frags = 1;
281 __skb_queue_tail(&np->rx_batch, skb);
282 }
283
284 /* Is the batch large enough to be worthwhile? */
285 if (i < (np->rx_target/2)) {
286 if (req_prod > np->rx.sring->req_prod)
287 goto push;
288 return;
289 }
290
291 /* Adjust our fill target if we risked running out of buffers. */
292 if (((req_prod - np->rx.sring->rsp_prod) < (np->rx_target / 4)) &&
293 ((np->rx_target *= 2) > np->rx_max_target))
294 np->rx_target = np->rx_max_target;
295
296 refill:
297 for (i = 0; ; i++) {
298 skb = __skb_dequeue(&np->rx_batch);
299 if (skb == NULL)
300 break;
301
302 skb->dev = dev;
303
304 id = xennet_rxidx(req_prod + i);
305
306 BUG_ON(np->rx_skbs[id]);
307 np->rx_skbs[id] = skb;
308
309 ref = gnttab_claim_grant_reference(&np->gref_rx_head);
310 BUG_ON((signed short)ref < 0);
311 np->grant_rx_ref[id] = ref;
312
313 pfn = page_to_pfn(skb_frag_page(&skb_shinfo(skb)->frags[0]));
314 vaddr = page_address(skb_frag_page(&skb_shinfo(skb)->frags[0]));
315
316 req = RING_GET_REQUEST(&np->rx, req_prod + i);
317 gnttab_grant_foreign_access_ref(ref,
318 np->xbdev->otherend_id,
319 pfn_to_mfn(pfn),
320 0);
321
322 req->id = id;
323 req->gref = ref;
324 }
325
326 wmb(); /* barrier so backend seens requests */
327
328 /* Above is a suitable barrier to ensure backend will see requests. */
329 np->rx.req_prod_pvt = req_prod + i;
330 push:
331 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&np->rx, notify);
332 if (notify)
333 notify_remote_via_irq(np->netdev->irq);
334}
335
336static int xennet_open(struct net_device *dev)
337{
338 struct netfront_info *np = netdev_priv(dev);
339
340 napi_enable(&np->napi);
341
342 spin_lock_bh(&np->rx_lock);
343 if (netif_carrier_ok(dev)) {
344 xennet_alloc_rx_buffers(dev);
345 np->rx.sring->rsp_event = np->rx.rsp_cons + 1;
346 if (RING_HAS_UNCONSUMED_RESPONSES(&np->rx))
347 napi_schedule(&np->napi);
348 }
349 spin_unlock_bh(&np->rx_lock);
350
351 netif_start_queue(dev);
352
353 return 0;
354}
355
356static void xennet_tx_buf_gc(struct net_device *dev)
357{
358 RING_IDX cons, prod;
359 unsigned short id;
360 struct netfront_info *np = netdev_priv(dev);
361 struct sk_buff *skb;
362
363 BUG_ON(!netif_carrier_ok(dev));
364
365 do {
366 prod = np->tx.sring->rsp_prod;
367 rmb(); /* Ensure we see responses up to 'rp'. */
368
369 for (cons = np->tx.rsp_cons; cons != prod; cons++) {
370 struct xen_netif_tx_response *txrsp;
371
372 txrsp = RING_GET_RESPONSE(&np->tx, cons);
373 if (txrsp->status == XEN_NETIF_RSP_NULL)
374 continue;
375
376 id = txrsp->id;
377 skb = np->tx_skbs[id].skb;
378 if (unlikely(gnttab_query_foreign_access(
379 np->grant_tx_ref[id]) != 0)) {
380 printk(KERN_ALERT "xennet_tx_buf_gc: warning "
381 "-- grant still in use by backend "
382 "domain.\n");
383 BUG();
384 }
385 gnttab_end_foreign_access_ref(
386 np->grant_tx_ref[id], GNTMAP_readonly);
387 gnttab_release_grant_reference(
388 &np->gref_tx_head, np->grant_tx_ref[id]);
389 np->grant_tx_ref[id] = GRANT_INVALID_REF;
390 add_id_to_freelist(&np->tx_skb_freelist, np->tx_skbs, id);
391 dev_kfree_skb_irq(skb);
392 }
393
394 np->tx.rsp_cons = prod;
395
396 /*
397 * Set a new event, then check for race with update of tx_cons.
398 * Note that it is essential to schedule a callback, no matter
399 * how few buffers are pending. Even if there is space in the
400 * transmit ring, higher layers may be blocked because too much
401 * data is outstanding: in such cases notification from Xen is
402 * likely to be the only kick that we'll get.
403 */
404 np->tx.sring->rsp_event =
405 prod + ((np->tx.sring->req_prod - prod) >> 1) + 1;
406 mb(); /* update shared area */
407 } while ((cons == prod) && (prod != np->tx.sring->rsp_prod));
408
409 xennet_maybe_wake_tx(dev);
410}
411
412static void xennet_make_frags(struct sk_buff *skb, struct net_device *dev,
413 struct xen_netif_tx_request *tx)
414{
415 struct netfront_info *np = netdev_priv(dev);
416 char *data = skb->data;
417 unsigned long mfn;
418 RING_IDX prod = np->tx.req_prod_pvt;
419 int frags = skb_shinfo(skb)->nr_frags;
420 unsigned int offset = offset_in_page(data);
421 unsigned int len = skb_headlen(skb);
422 unsigned int id;
423 grant_ref_t ref;
424 int i;
425
426 /* While the header overlaps a page boundary (including being
427 larger than a page), split it it into page-sized chunks. */
428 while (len > PAGE_SIZE - offset) {
429 tx->size = PAGE_SIZE - offset;
430 tx->flags |= XEN_NETTXF_more_data;
431 len -= tx->size;
432 data += tx->size;
433 offset = 0;
434
435 id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs);
436 np->tx_skbs[id].skb = skb_get(skb);
437 tx = RING_GET_REQUEST(&np->tx, prod++);
438 tx->id = id;
439 ref = gnttab_claim_grant_reference(&np->gref_tx_head);
440 BUG_ON((signed short)ref < 0);
441
442 mfn = virt_to_mfn(data);
443 gnttab_grant_foreign_access_ref(ref, np->xbdev->otherend_id,
444 mfn, GNTMAP_readonly);
445
446 tx->gref = np->grant_tx_ref[id] = ref;
447 tx->offset = offset;
448 tx->size = len;
449 tx->flags = 0;
450 }
451
452 /* Grant backend access to each skb fragment page. */
453 for (i = 0; i < frags; i++) {
454 skb_frag_t *frag = skb_shinfo(skb)->frags + i;
455
456 tx->flags |= XEN_NETTXF_more_data;
457
458 id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs);
459 np->tx_skbs[id].skb = skb_get(skb);
460 tx = RING_GET_REQUEST(&np->tx, prod++);
461 tx->id = id;
462 ref = gnttab_claim_grant_reference(&np->gref_tx_head);
463 BUG_ON((signed short)ref < 0);
464
465 mfn = pfn_to_mfn(page_to_pfn(skb_frag_page(frag)));
466 gnttab_grant_foreign_access_ref(ref, np->xbdev->otherend_id,
467 mfn, GNTMAP_readonly);
468
469 tx->gref = np->grant_tx_ref[id] = ref;
470 tx->offset = frag->page_offset;
471 tx->size = skb_frag_size(frag);
472 tx->flags = 0;
473 }
474
475 np->tx.req_prod_pvt = prod;
476}
477
478static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev)
479{
480 unsigned short id;
481 struct netfront_info *np = netdev_priv(dev);
482 struct netfront_stats *stats = this_cpu_ptr(np->stats);
483 struct xen_netif_tx_request *tx;
484 struct xen_netif_extra_info *extra;
485 char *data = skb->data;
486 RING_IDX i;
487 grant_ref_t ref;
488 unsigned long mfn;
489 int notify;
490 int frags = skb_shinfo(skb)->nr_frags;
491 unsigned int offset = offset_in_page(data);
492 unsigned int len = skb_headlen(skb);
493 unsigned long flags;
494
495 frags += DIV_ROUND_UP(offset + len, PAGE_SIZE);
496 if (unlikely(frags > MAX_SKB_FRAGS + 1)) {
497 printk(KERN_ALERT "xennet: skb rides the rocket: %d frags\n",
498 frags);
499 dump_stack();
500 goto drop;
501 }
502
503 spin_lock_irqsave(&np->tx_lock, flags);
504
505 if (unlikely(!netif_carrier_ok(dev) ||
506 (frags > 1 && !xennet_can_sg(dev)) ||
507 netif_needs_gso(skb, netif_skb_features(skb)))) {
508 spin_unlock_irqrestore(&np->tx_lock, flags);
509 goto drop;
510 }
511
512 i = np->tx.req_prod_pvt;
513
514 id = get_id_from_freelist(&np->tx_skb_freelist, np->tx_skbs);
515 np->tx_skbs[id].skb = skb;
516
517 tx = RING_GET_REQUEST(&np->tx, i);
518
519 tx->id = id;
520 ref = gnttab_claim_grant_reference(&np->gref_tx_head);
521 BUG_ON((signed short)ref < 0);
522 mfn = virt_to_mfn(data);
523 gnttab_grant_foreign_access_ref(
524 ref, np->xbdev->otherend_id, mfn, GNTMAP_readonly);
525 tx->gref = np->grant_tx_ref[id] = ref;
526 tx->offset = offset;
527 tx->size = len;
528 extra = NULL;
529
530 tx->flags = 0;
531 if (skb->ip_summed == CHECKSUM_PARTIAL)
532 /* local packet? */
533 tx->flags |= XEN_NETTXF_csum_blank | XEN_NETTXF_data_validated;
534 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
535 /* remote but checksummed. */
536 tx->flags |= XEN_NETTXF_data_validated;
537
538 if (skb_shinfo(skb)->gso_size) {
539 struct xen_netif_extra_info *gso;
540
541 gso = (struct xen_netif_extra_info *)
542 RING_GET_REQUEST(&np->tx, ++i);
543
544 if (extra)
545 extra->flags |= XEN_NETIF_EXTRA_FLAG_MORE;
546 else
547 tx->flags |= XEN_NETTXF_extra_info;
548
549 gso->u.gso.size = skb_shinfo(skb)->gso_size;
550 gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
551 gso->u.gso.pad = 0;
552 gso->u.gso.features = 0;
553
554 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
555 gso->flags = 0;
556 extra = gso;
557 }
558
559 np->tx.req_prod_pvt = i + 1;
560
561 xennet_make_frags(skb, dev, tx);
562 tx->size = skb->len;
563
564 RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&np->tx, notify);
565 if (notify)
566 notify_remote_via_irq(np->netdev->irq);
567
568 u64_stats_update_begin(&stats->syncp);
569 stats->tx_bytes += skb->len;
570 stats->tx_packets++;
571 u64_stats_update_end(&stats->syncp);
572
573 /* Note: It is not safe to access skb after xennet_tx_buf_gc()! */
574 xennet_tx_buf_gc(dev);
575
576 if (!netfront_tx_slot_available(np))
577 netif_stop_queue(dev);
578
579 spin_unlock_irqrestore(&np->tx_lock, flags);
580
581 return NETDEV_TX_OK;
582
583 drop:
584 dev->stats.tx_dropped++;
585 dev_kfree_skb(skb);
586 return NETDEV_TX_OK;
587}
588
589static int xennet_close(struct net_device *dev)
590{
591 struct netfront_info *np = netdev_priv(dev);
592 netif_stop_queue(np->netdev);
593 napi_disable(&np->napi);
594 return 0;
595}
596
597static void xennet_move_rx_slot(struct netfront_info *np, struct sk_buff *skb,
598 grant_ref_t ref)
599{
600 int new = xennet_rxidx(np->rx.req_prod_pvt);
601
602 BUG_ON(np->rx_skbs[new]);
603 np->rx_skbs[new] = skb;
604 np->grant_rx_ref[new] = ref;
605 RING_GET_REQUEST(&np->rx, np->rx.req_prod_pvt)->id = new;
606 RING_GET_REQUEST(&np->rx, np->rx.req_prod_pvt)->gref = ref;
607 np->rx.req_prod_pvt++;
608}
609
610static int xennet_get_extras(struct netfront_info *np,
611 struct xen_netif_extra_info *extras,
612 RING_IDX rp)
613
614{
615 struct xen_netif_extra_info *extra;
616 struct device *dev = &np->netdev->dev;
617 RING_IDX cons = np->rx.rsp_cons;
618 int err = 0;
619
620 do {
621 struct sk_buff *skb;
622 grant_ref_t ref;
623
624 if (unlikely(cons + 1 == rp)) {
625 if (net_ratelimit())
626 dev_warn(dev, "Missing extra info\n");
627 err = -EBADR;
628 break;
629 }
630
631 extra = (struct xen_netif_extra_info *)
632 RING_GET_RESPONSE(&np->rx, ++cons);
633
634 if (unlikely(!extra->type ||
635 extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
636 if (net_ratelimit())
637 dev_warn(dev, "Invalid extra type: %d\n",
638 extra->type);
639 err = -EINVAL;
640 } else {
641 memcpy(&extras[extra->type - 1], extra,
642 sizeof(*extra));
643 }
644
645 skb = xennet_get_rx_skb(np, cons);
646 ref = xennet_get_rx_ref(np, cons);
647 xennet_move_rx_slot(np, skb, ref);
648 } while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
649
650 np->rx.rsp_cons = cons;
651 return err;
652}
653
654static int xennet_get_responses(struct netfront_info *np,
655 struct netfront_rx_info *rinfo, RING_IDX rp,
656 struct sk_buff_head *list)
657{
658 struct xen_netif_rx_response *rx = &rinfo->rx;
659 struct xen_netif_extra_info *extras = rinfo->extras;
660 struct device *dev = &np->netdev->dev;
661 RING_IDX cons = np->rx.rsp_cons;
662 struct sk_buff *skb = xennet_get_rx_skb(np, cons);
663 grant_ref_t ref = xennet_get_rx_ref(np, cons);
664 int max = MAX_SKB_FRAGS + (rx->status <= RX_COPY_THRESHOLD);
665 int frags = 1;
666 int err = 0;
667 unsigned long ret;
668
669 if (rx->flags & XEN_NETRXF_extra_info) {
670 err = xennet_get_extras(np, extras, rp);
671 cons = np->rx.rsp_cons;
672 }
673
674 for (;;) {
675 if (unlikely(rx->status < 0 ||
676 rx->offset + rx->status > PAGE_SIZE)) {
677 if (net_ratelimit())
678 dev_warn(dev, "rx->offset: %x, size: %u\n",
679 rx->offset, rx->status);
680 xennet_move_rx_slot(np, skb, ref);
681 err = -EINVAL;
682 goto next;
683 }
684
685 /*
686 * This definitely indicates a bug, either in this driver or in
687 * the backend driver. In future this should flag the bad
688 * situation to the system controller to reboot the backed.
689 */
690 if (ref == GRANT_INVALID_REF) {
691 if (net_ratelimit())
692 dev_warn(dev, "Bad rx response id %d.\n",
693 rx->id);
694 err = -EINVAL;
695 goto next;
696 }
697
698 ret = gnttab_end_foreign_access_ref(ref, 0);
699 BUG_ON(!ret);
700
701 gnttab_release_grant_reference(&np->gref_rx_head, ref);
702
703 __skb_queue_tail(list, skb);
704
705next:
706 if (!(rx->flags & XEN_NETRXF_more_data))
707 break;
708
709 if (cons + frags == rp) {
710 if (net_ratelimit())
711 dev_warn(dev, "Need more frags\n");
712 err = -ENOENT;
713 break;
714 }
715
716 rx = RING_GET_RESPONSE(&np->rx, cons + frags);
717 skb = xennet_get_rx_skb(np, cons + frags);
718 ref = xennet_get_rx_ref(np, cons + frags);
719 frags++;
720 }
721
722 if (unlikely(frags > max)) {
723 if (net_ratelimit())
724 dev_warn(dev, "Too many frags\n");
725 err = -E2BIG;
726 }
727
728 if (unlikely(err))
729 np->rx.rsp_cons = cons + frags;
730
731 return err;
732}
733
734static int xennet_set_skb_gso(struct sk_buff *skb,
735 struct xen_netif_extra_info *gso)
736{
737 if (!gso->u.gso.size) {
738 if (net_ratelimit())
739 printk(KERN_WARNING "GSO size must not be zero.\n");
740 return -EINVAL;
741 }
742
743 /* Currently only TCPv4 S.O. is supported. */
744 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
745 if (net_ratelimit())
746 printk(KERN_WARNING "Bad GSO type %d.\n", gso->u.gso.type);
747 return -EINVAL;
748 }
749
750 skb_shinfo(skb)->gso_size = gso->u.gso.size;
751 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
752
753 /* Header must be checked, and gso_segs computed. */
754 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
755 skb_shinfo(skb)->gso_segs = 0;
756
757 return 0;
758}
759
760static RING_IDX xennet_fill_frags(struct netfront_info *np,
761 struct sk_buff *skb,
762 struct sk_buff_head *list)
763{
764 struct skb_shared_info *shinfo = skb_shinfo(skb);
765 int nr_frags = shinfo->nr_frags;
766 RING_IDX cons = np->rx.rsp_cons;
767 struct sk_buff *nskb;
768
769 while ((nskb = __skb_dequeue(list))) {
770 struct xen_netif_rx_response *rx =
771 RING_GET_RESPONSE(&np->rx, ++cons);
772 skb_frag_t *nfrag = &skb_shinfo(nskb)->frags[0];
773
774 __skb_fill_page_desc(skb, nr_frags,
775 skb_frag_page(nfrag),
776 rx->offset, rx->status);
777
778 skb->data_len += rx->status;
779
780 skb_shinfo(nskb)->nr_frags = 0;
781 kfree_skb(nskb);
782
783 nr_frags++;
784 }
785
786 shinfo->nr_frags = nr_frags;
787 return cons;
788}
789
790static int checksum_setup(struct net_device *dev, struct sk_buff *skb)
791{
792 struct iphdr *iph;
793 unsigned char *th;
794 int err = -EPROTO;
795 int recalculate_partial_csum = 0;
796
797 /*
798 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
799 * peers can fail to set NETRXF_csum_blank when sending a GSO
800 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
801 * recalculate the partial checksum.
802 */
803 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
804 struct netfront_info *np = netdev_priv(dev);
805 np->rx_gso_checksum_fixup++;
806 skb->ip_summed = CHECKSUM_PARTIAL;
807 recalculate_partial_csum = 1;
808 }
809
810 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
811 if (skb->ip_summed != CHECKSUM_PARTIAL)
812 return 0;
813
814 if (skb->protocol != htons(ETH_P_IP))
815 goto out;
816
817 iph = (void *)skb->data;
818 th = skb->data + 4 * iph->ihl;
819 if (th >= skb_tail_pointer(skb))
820 goto out;
821
822 skb->csum_start = th - skb->head;
823 switch (iph->protocol) {
824 case IPPROTO_TCP:
825 skb->csum_offset = offsetof(struct tcphdr, check);
826
827 if (recalculate_partial_csum) {
828 struct tcphdr *tcph = (struct tcphdr *)th;
829 tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
830 skb->len - iph->ihl*4,
831 IPPROTO_TCP, 0);
832 }
833 break;
834 case IPPROTO_UDP:
835 skb->csum_offset = offsetof(struct udphdr, check);
836
837 if (recalculate_partial_csum) {
838 struct udphdr *udph = (struct udphdr *)th;
839 udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
840 skb->len - iph->ihl*4,
841 IPPROTO_UDP, 0);
842 }
843 break;
844 default:
845 if (net_ratelimit())
846 printk(KERN_ERR "Attempting to checksum a non-"
847 "TCP/UDP packet, dropping a protocol"
848 " %d packet", iph->protocol);
849 goto out;
850 }
851
852 if ((th + skb->csum_offset + 2) > skb_tail_pointer(skb))
853 goto out;
854
855 err = 0;
856
857out:
858 return err;
859}
860
861static int handle_incoming_queue(struct net_device *dev,
862 struct sk_buff_head *rxq)
863{
864 struct netfront_info *np = netdev_priv(dev);
865 struct netfront_stats *stats = this_cpu_ptr(np->stats);
866 int packets_dropped = 0;
867 struct sk_buff *skb;
868
869 while ((skb = __skb_dequeue(rxq)) != NULL) {
870 struct page *page = NETFRONT_SKB_CB(skb)->page;
871 void *vaddr = page_address(page);
872 unsigned offset = NETFRONT_SKB_CB(skb)->offset;
873
874 memcpy(skb->data, vaddr + offset,
875 skb_headlen(skb));
876
877 if (page != skb_frag_page(&skb_shinfo(skb)->frags[0]))
878 __free_page(page);
879
880 /* Ethernet work: Delayed to here as it peeks the header. */
881 skb->protocol = eth_type_trans(skb, dev);
882
883 if (checksum_setup(dev, skb)) {
884 kfree_skb(skb);
885 packets_dropped++;
886 dev->stats.rx_errors++;
887 continue;
888 }
889
890 u64_stats_update_begin(&stats->syncp);
891 stats->rx_packets++;
892 stats->rx_bytes += skb->len;
893 u64_stats_update_end(&stats->syncp);
894
895 /* Pass it up. */
896 netif_receive_skb(skb);
897 }
898
899 return packets_dropped;
900}
901
902static int xennet_poll(struct napi_struct *napi, int budget)
903{
904 struct netfront_info *np = container_of(napi, struct netfront_info, napi);
905 struct net_device *dev = np->netdev;
906 struct sk_buff *skb;
907 struct netfront_rx_info rinfo;
908 struct xen_netif_rx_response *rx = &rinfo.rx;
909 struct xen_netif_extra_info *extras = rinfo.extras;
910 RING_IDX i, rp;
911 int work_done;
912 struct sk_buff_head rxq;
913 struct sk_buff_head errq;
914 struct sk_buff_head tmpq;
915 unsigned long flags;
916 unsigned int len;
917 int err;
918
919 spin_lock(&np->rx_lock);
920
921 skb_queue_head_init(&rxq);
922 skb_queue_head_init(&errq);
923 skb_queue_head_init(&tmpq);
924
925 rp = np->rx.sring->rsp_prod;
926 rmb(); /* Ensure we see queued responses up to 'rp'. */
927
928 i = np->rx.rsp_cons;
929 work_done = 0;
930 while ((i != rp) && (work_done < budget)) {
931 memcpy(rx, RING_GET_RESPONSE(&np->rx, i), sizeof(*rx));
932 memset(extras, 0, sizeof(rinfo.extras));
933
934 err = xennet_get_responses(np, &rinfo, rp, &tmpq);
935
936 if (unlikely(err)) {
937err:
938 while ((skb = __skb_dequeue(&tmpq)))
939 __skb_queue_tail(&errq, skb);
940 dev->stats.rx_errors++;
941 i = np->rx.rsp_cons;
942 continue;
943 }
944
945 skb = __skb_dequeue(&tmpq);
946
947 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
948 struct xen_netif_extra_info *gso;
949 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
950
951 if (unlikely(xennet_set_skb_gso(skb, gso))) {
952 __skb_queue_head(&tmpq, skb);
953 np->rx.rsp_cons += skb_queue_len(&tmpq);
954 goto err;
955 }
956 }
957
958 NETFRONT_SKB_CB(skb)->page =
959 skb_frag_page(&skb_shinfo(skb)->frags[0]);
960 NETFRONT_SKB_CB(skb)->offset = rx->offset;
961
962 len = rx->status;
963 if (len > RX_COPY_THRESHOLD)
964 len = RX_COPY_THRESHOLD;
965 skb_put(skb, len);
966
967 if (rx->status > len) {
968 skb_shinfo(skb)->frags[0].page_offset =
969 rx->offset + len;
970 skb_frag_size_set(&skb_shinfo(skb)->frags[0], rx->status - len);
971 skb->data_len = rx->status - len;
972 } else {
973 __skb_fill_page_desc(skb, 0, NULL, 0, 0);
974 skb_shinfo(skb)->nr_frags = 0;
975 }
976
977 i = xennet_fill_frags(np, skb, &tmpq);
978
979 /*
980 * Truesize approximates the size of true data plus
981 * any supervisor overheads. Adding hypervisor
982 * overheads has been shown to significantly reduce
983 * achievable bandwidth with the default receive
984 * buffer size. It is therefore not wise to account
985 * for it here.
986 *
987 * After alloc_skb(RX_COPY_THRESHOLD), truesize is set
988 * to RX_COPY_THRESHOLD + the supervisor
989 * overheads. Here, we add the size of the data pulled
990 * in xennet_fill_frags().
991 *
992 * We also adjust for any unused space in the main
993 * data area by subtracting (RX_COPY_THRESHOLD -
994 * len). This is especially important with drivers
995 * which split incoming packets into header and data,
996 * using only 66 bytes of the main data area (see the
997 * e1000 driver for example.) On such systems,
998 * without this last adjustement, our achievable
999 * receive throughout using the standard receive
1000 * buffer size was cut by 25%(!!!).
1001 */
1002 skb->truesize += skb->data_len - (RX_COPY_THRESHOLD - len);
1003 skb->len += skb->data_len;
1004
1005 if (rx->flags & XEN_NETRXF_csum_blank)
1006 skb->ip_summed = CHECKSUM_PARTIAL;
1007 else if (rx->flags & XEN_NETRXF_data_validated)
1008 skb->ip_summed = CHECKSUM_UNNECESSARY;
1009
1010 __skb_queue_tail(&rxq, skb);
1011
1012 np->rx.rsp_cons = ++i;
1013 work_done++;
1014 }
1015
1016 __skb_queue_purge(&errq);
1017
1018 work_done -= handle_incoming_queue(dev, &rxq);
1019
1020 /* If we get a callback with very few responses, reduce fill target. */
1021 /* NB. Note exponential increase, linear decrease. */
1022 if (((np->rx.req_prod_pvt - np->rx.sring->rsp_prod) >
1023 ((3*np->rx_target) / 4)) &&
1024 (--np->rx_target < np->rx_min_target))
1025 np->rx_target = np->rx_min_target;
1026
1027 xennet_alloc_rx_buffers(dev);
1028
1029 if (work_done < budget) {
1030 int more_to_do = 0;
1031
1032 local_irq_save(flags);
1033
1034 RING_FINAL_CHECK_FOR_RESPONSES(&np->rx, more_to_do);
1035 if (!more_to_do)
1036 __napi_complete(napi);
1037
1038 local_irq_restore(flags);
1039 }
1040
1041 spin_unlock(&np->rx_lock);
1042
1043 return work_done;
1044}
1045
1046static int xennet_change_mtu(struct net_device *dev, int mtu)
1047{
1048 int max = xennet_can_sg(dev) ? 65535 - ETH_HLEN : ETH_DATA_LEN;
1049
1050 if (mtu > max)
1051 return -EINVAL;
1052 dev->mtu = mtu;
1053 return 0;
1054}
1055
1056static struct rtnl_link_stats64 *xennet_get_stats64(struct net_device *dev,
1057 struct rtnl_link_stats64 *tot)
1058{
1059 struct netfront_info *np = netdev_priv(dev);
1060 int cpu;
1061
1062 for_each_possible_cpu(cpu) {
1063 struct netfront_stats *stats = per_cpu_ptr(np->stats, cpu);
1064 u64 rx_packets, rx_bytes, tx_packets, tx_bytes;
1065 unsigned int start;
1066
1067 do {
1068 start = u64_stats_fetch_begin_bh(&stats->syncp);
1069
1070 rx_packets = stats->rx_packets;
1071 tx_packets = stats->tx_packets;
1072 rx_bytes = stats->rx_bytes;
1073 tx_bytes = stats->tx_bytes;
1074 } while (u64_stats_fetch_retry_bh(&stats->syncp, start));
1075
1076 tot->rx_packets += rx_packets;
1077 tot->tx_packets += tx_packets;
1078 tot->rx_bytes += rx_bytes;
1079 tot->tx_bytes += tx_bytes;
1080 }
1081
1082 tot->rx_errors = dev->stats.rx_errors;
1083 tot->tx_dropped = dev->stats.tx_dropped;
1084
1085 return tot;
1086}
1087
1088static void xennet_release_tx_bufs(struct netfront_info *np)
1089{
1090 struct sk_buff *skb;
1091 int i;
1092
1093 for (i = 0; i < NET_TX_RING_SIZE; i++) {
1094 /* Skip over entries which are actually freelist references */
1095 if (skb_entry_is_link(&np->tx_skbs[i]))
1096 continue;
1097
1098 skb = np->tx_skbs[i].skb;
1099 gnttab_end_foreign_access_ref(np->grant_tx_ref[i],
1100 GNTMAP_readonly);
1101 gnttab_release_grant_reference(&np->gref_tx_head,
1102 np->grant_tx_ref[i]);
1103 np->grant_tx_ref[i] = GRANT_INVALID_REF;
1104 add_id_to_freelist(&np->tx_skb_freelist, np->tx_skbs, i);
1105 dev_kfree_skb_irq(skb);
1106 }
1107}
1108
1109static void xennet_release_rx_bufs(struct netfront_info *np)
1110{
1111 struct mmu_update *mmu = np->rx_mmu;
1112 struct multicall_entry *mcl = np->rx_mcl;
1113 struct sk_buff_head free_list;
1114 struct sk_buff *skb;
1115 unsigned long mfn;
1116 int xfer = 0, noxfer = 0, unused = 0;
1117 int id, ref;
1118
1119 dev_warn(&np->netdev->dev, "%s: fix me for copying receiver.\n",
1120 __func__);
1121 return;
1122
1123 skb_queue_head_init(&free_list);
1124
1125 spin_lock_bh(&np->rx_lock);
1126
1127 for (id = 0; id < NET_RX_RING_SIZE; id++) {
1128 ref = np->grant_rx_ref[id];
1129 if (ref == GRANT_INVALID_REF) {
1130 unused++;
1131 continue;
1132 }
1133
1134 skb = np->rx_skbs[id];
1135 mfn = gnttab_end_foreign_transfer_ref(ref);
1136 gnttab_release_grant_reference(&np->gref_rx_head, ref);
1137 np->grant_rx_ref[id] = GRANT_INVALID_REF;
1138
1139 if (0 == mfn) {
1140 skb_shinfo(skb)->nr_frags = 0;
1141 dev_kfree_skb(skb);
1142 noxfer++;
1143 continue;
1144 }
1145
1146 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1147 /* Remap the page. */
1148 const struct page *page =
1149 skb_frag_page(&skb_shinfo(skb)->frags[0]);
1150 unsigned long pfn = page_to_pfn(page);
1151 void *vaddr = page_address(page);
1152
1153 MULTI_update_va_mapping(mcl, (unsigned long)vaddr,
1154 mfn_pte(mfn, PAGE_KERNEL),
1155 0);
1156 mcl++;
1157 mmu->ptr = ((u64)mfn << PAGE_SHIFT)
1158 | MMU_MACHPHYS_UPDATE;
1159 mmu->val = pfn;
1160 mmu++;
1161
1162 set_phys_to_machine(pfn, mfn);
1163 }
1164 __skb_queue_tail(&free_list, skb);
1165 xfer++;
1166 }
1167
1168 dev_info(&np->netdev->dev, "%s: %d xfer, %d noxfer, %d unused\n",
1169 __func__, xfer, noxfer, unused);
1170
1171 if (xfer) {
1172 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
1173 /* Do all the remapping work and M2P updates. */
1174 MULTI_mmu_update(mcl, np->rx_mmu, mmu - np->rx_mmu,
1175 NULL, DOMID_SELF);
1176 mcl++;
1177 HYPERVISOR_multicall(np->rx_mcl, mcl - np->rx_mcl);
1178 }
1179 }
1180
1181 __skb_queue_purge(&free_list);
1182
1183 spin_unlock_bh(&np->rx_lock);
1184}
1185
1186static void xennet_uninit(struct net_device *dev)
1187{
1188 struct netfront_info *np = netdev_priv(dev);
1189 xennet_release_tx_bufs(np);
1190 xennet_release_rx_bufs(np);
1191 gnttab_free_grant_references(np->gref_tx_head);
1192 gnttab_free_grant_references(np->gref_rx_head);
1193}
1194
1195static netdev_features_t xennet_fix_features(struct net_device *dev,
1196 netdev_features_t features)
1197{
1198 struct netfront_info *np = netdev_priv(dev);
1199 int val;
1200
1201 if (features & NETIF_F_SG) {
1202 if (xenbus_scanf(XBT_NIL, np->xbdev->otherend, "feature-sg",
1203 "%d", &val) < 0)
1204 val = 0;
1205
1206 if (!val)
1207 features &= ~NETIF_F_SG;
1208 }
1209
1210 if (features & NETIF_F_TSO) {
1211 if (xenbus_scanf(XBT_NIL, np->xbdev->otherend,
1212 "feature-gso-tcpv4", "%d", &val) < 0)
1213 val = 0;
1214
1215 if (!val)
1216 features &= ~NETIF_F_TSO;
1217 }
1218
1219 return features;
1220}
1221
1222static int xennet_set_features(struct net_device *dev,
1223 netdev_features_t features)
1224{
1225 if (!(features & NETIF_F_SG) && dev->mtu > ETH_DATA_LEN) {
1226 netdev_info(dev, "Reducing MTU because no SG offload");
1227 dev->mtu = ETH_DATA_LEN;
1228 }
1229
1230 return 0;
1231}
1232
1233static irqreturn_t xennet_interrupt(int irq, void *dev_id)
1234{
1235 struct net_device *dev = dev_id;
1236 struct netfront_info *np = netdev_priv(dev);
1237 unsigned long flags;
1238
1239 spin_lock_irqsave(&np->tx_lock, flags);
1240
1241 if (likely(netif_carrier_ok(dev))) {
1242 xennet_tx_buf_gc(dev);
1243 /* Under tx_lock: protects access to rx shared-ring indexes. */
1244 if (RING_HAS_UNCONSUMED_RESPONSES(&np->rx))
1245 napi_schedule(&np->napi);
1246 }
1247
1248 spin_unlock_irqrestore(&np->tx_lock, flags);
1249
1250 return IRQ_HANDLED;
1251}
1252
1253#ifdef CONFIG_NET_POLL_CONTROLLER
1254static void xennet_poll_controller(struct net_device *dev)
1255{
1256 xennet_interrupt(0, dev);
1257}
1258#endif
1259
1260static const struct net_device_ops xennet_netdev_ops = {
1261 .ndo_open = xennet_open,
1262 .ndo_uninit = xennet_uninit,
1263 .ndo_stop = xennet_close,
1264 .ndo_start_xmit = xennet_start_xmit,
1265 .ndo_change_mtu = xennet_change_mtu,
1266 .ndo_get_stats64 = xennet_get_stats64,
1267 .ndo_set_mac_address = eth_mac_addr,
1268 .ndo_validate_addr = eth_validate_addr,
1269 .ndo_fix_features = xennet_fix_features,
1270 .ndo_set_features = xennet_set_features,
1271#ifdef CONFIG_NET_POLL_CONTROLLER
1272 .ndo_poll_controller = xennet_poll_controller,
1273#endif
1274};
1275
1276static struct net_device * __devinit xennet_create_dev(struct xenbus_device *dev)
1277{
1278 int i, err;
1279 struct net_device *netdev;
1280 struct netfront_info *np;
1281
1282 netdev = alloc_etherdev(sizeof(struct netfront_info));
1283 if (!netdev)
1284 return ERR_PTR(-ENOMEM);
1285
1286 np = netdev_priv(netdev);
1287 np->xbdev = dev;
1288
1289 spin_lock_init(&np->tx_lock);
1290 spin_lock_init(&np->rx_lock);
1291
1292 skb_queue_head_init(&np->rx_batch);
1293 np->rx_target = RX_DFL_MIN_TARGET;
1294 np->rx_min_target = RX_DFL_MIN_TARGET;
1295 np->rx_max_target = RX_MAX_TARGET;
1296
1297 init_timer(&np->rx_refill_timer);
1298 np->rx_refill_timer.data = (unsigned long)netdev;
1299 np->rx_refill_timer.function = rx_refill_timeout;
1300
1301 err = -ENOMEM;
1302 np->stats = alloc_percpu(struct netfront_stats);
1303 if (np->stats == NULL)
1304 goto exit;
1305
1306 /* Initialise tx_skbs as a free chain containing every entry. */
1307 np->tx_skb_freelist = 0;
1308 for (i = 0; i < NET_TX_RING_SIZE; i++) {
1309 skb_entry_set_link(&np->tx_skbs[i], i+1);
1310 np->grant_tx_ref[i] = GRANT_INVALID_REF;
1311 }
1312
1313 /* Clear out rx_skbs */
1314 for (i = 0; i < NET_RX_RING_SIZE; i++) {
1315 np->rx_skbs[i] = NULL;
1316 np->grant_rx_ref[i] = GRANT_INVALID_REF;
1317 }
1318
1319 /* A grant for every tx ring slot */
1320 if (gnttab_alloc_grant_references(TX_MAX_TARGET,
1321 &np->gref_tx_head) < 0) {
1322 printk(KERN_ALERT "#### netfront can't alloc tx grant refs\n");
1323 err = -ENOMEM;
1324 goto exit_free_stats;
1325 }
1326 /* A grant for every rx ring slot */
1327 if (gnttab_alloc_grant_references(RX_MAX_TARGET,
1328 &np->gref_rx_head) < 0) {
1329 printk(KERN_ALERT "#### netfront can't alloc rx grant refs\n");
1330 err = -ENOMEM;
1331 goto exit_free_tx;
1332 }
1333
1334 netdev->netdev_ops = &xennet_netdev_ops;
1335
1336 netif_napi_add(netdev, &np->napi, xennet_poll, 64);
1337 netdev->features = NETIF_F_IP_CSUM | NETIF_F_RXCSUM |
1338 NETIF_F_GSO_ROBUST;
1339 netdev->hw_features = NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO;
1340
1341 /*
1342 * Assume that all hw features are available for now. This set
1343 * will be adjusted by the call to netdev_update_features() in
1344 * xennet_connect() which is the earliest point where we can
1345 * negotiate with the backend regarding supported features.
1346 */
1347 netdev->features |= netdev->hw_features;
1348
1349 SET_ETHTOOL_OPS(netdev, &xennet_ethtool_ops);
1350 SET_NETDEV_DEV(netdev, &dev->dev);
1351
1352 np->netdev = netdev;
1353
1354 netif_carrier_off(netdev);
1355
1356 return netdev;
1357
1358 exit_free_tx:
1359 gnttab_free_grant_references(np->gref_tx_head);
1360 exit_free_stats:
1361 free_percpu(np->stats);
1362 exit:
1363 free_netdev(netdev);
1364 return ERR_PTR(err);
1365}
1366
1367/**
1368 * Entry point to this code when a new device is created. Allocate the basic
1369 * structures and the ring buffers for communication with the backend, and
1370 * inform the backend of the appropriate details for those.
1371 */
1372static int __devinit netfront_probe(struct xenbus_device *dev,
1373 const struct xenbus_device_id *id)
1374{
1375 int err;
1376 struct net_device *netdev;
1377 struct netfront_info *info;
1378
1379 netdev = xennet_create_dev(dev);
1380 if (IS_ERR(netdev)) {
1381 err = PTR_ERR(netdev);
1382 xenbus_dev_fatal(dev, err, "creating netdev");
1383 return err;
1384 }
1385
1386 info = netdev_priv(netdev);
1387 dev_set_drvdata(&dev->dev, info);
1388
1389 err = register_netdev(info->netdev);
1390 if (err) {
1391 printk(KERN_WARNING "%s: register_netdev err=%d\n",
1392 __func__, err);
1393 goto fail;
1394 }
1395
1396 err = xennet_sysfs_addif(info->netdev);
1397 if (err) {
1398 unregister_netdev(info->netdev);
1399 printk(KERN_WARNING "%s: add sysfs failed err=%d\n",
1400 __func__, err);
1401 goto fail;
1402 }
1403
1404 return 0;
1405
1406 fail:
1407 free_netdev(netdev);
1408 dev_set_drvdata(&dev->dev, NULL);
1409 return err;
1410}
1411
1412static void xennet_end_access(int ref, void *page)
1413{
1414 /* This frees the page as a side-effect */
1415 if (ref != GRANT_INVALID_REF)
1416 gnttab_end_foreign_access(ref, 0, (unsigned long)page);
1417}
1418
1419static void xennet_disconnect_backend(struct netfront_info *info)
1420{
1421 /* Stop old i/f to prevent errors whilst we rebuild the state. */
1422 spin_lock_bh(&info->rx_lock);
1423 spin_lock_irq(&info->tx_lock);
1424 netif_carrier_off(info->netdev);
1425 spin_unlock_irq(&info->tx_lock);
1426 spin_unlock_bh(&info->rx_lock);
1427
1428 if (info->netdev->irq)
1429 unbind_from_irqhandler(info->netdev->irq, info->netdev);
1430 info->evtchn = info->netdev->irq = 0;
1431
1432 /* End access and free the pages */
1433 xennet_end_access(info->tx_ring_ref, info->tx.sring);
1434 xennet_end_access(info->rx_ring_ref, info->rx.sring);
1435
1436 info->tx_ring_ref = GRANT_INVALID_REF;
1437 info->rx_ring_ref = GRANT_INVALID_REF;
1438 info->tx.sring = NULL;
1439 info->rx.sring = NULL;
1440}
1441
1442/**
1443 * We are reconnecting to the backend, due to a suspend/resume, or a backend
1444 * driver restart. We tear down our netif structure and recreate it, but
1445 * leave the device-layer structures intact so that this is transparent to the
1446 * rest of the kernel.
1447 */
1448static int netfront_resume(struct xenbus_device *dev)
1449{
1450 struct netfront_info *info = dev_get_drvdata(&dev->dev);
1451
1452 dev_dbg(&dev->dev, "%s\n", dev->nodename);
1453
1454 xennet_disconnect_backend(info);
1455 return 0;
1456}
1457
1458static int xen_net_read_mac(struct xenbus_device *dev, u8 mac[])
1459{
1460 char *s, *e, *macstr;
1461 int i;
1462
1463 macstr = s = xenbus_read(XBT_NIL, dev->nodename, "mac", NULL);
1464 if (IS_ERR(macstr))
1465 return PTR_ERR(macstr);
1466
1467 for (i = 0; i < ETH_ALEN; i++) {
1468 mac[i] = simple_strtoul(s, &e, 16);
1469 if ((s == e) || (*e != ((i == ETH_ALEN-1) ? '\0' : ':'))) {
1470 kfree(macstr);
1471 return -ENOENT;
1472 }
1473 s = e+1;
1474 }
1475
1476 kfree(macstr);
1477 return 0;
1478}
1479
1480static int setup_netfront(struct xenbus_device *dev, struct netfront_info *info)
1481{
1482 struct xen_netif_tx_sring *txs;
1483 struct xen_netif_rx_sring *rxs;
1484 int err;
1485 struct net_device *netdev = info->netdev;
1486
1487 info->tx_ring_ref = GRANT_INVALID_REF;
1488 info->rx_ring_ref = GRANT_INVALID_REF;
1489 info->rx.sring = NULL;
1490 info->tx.sring = NULL;
1491 netdev->irq = 0;
1492
1493 err = xen_net_read_mac(dev, netdev->dev_addr);
1494 if (err) {
1495 xenbus_dev_fatal(dev, err, "parsing %s/mac", dev->nodename);
1496 goto fail;
1497 }
1498
1499 txs = (struct xen_netif_tx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH);
1500 if (!txs) {
1501 err = -ENOMEM;
1502 xenbus_dev_fatal(dev, err, "allocating tx ring page");
1503 goto fail;
1504 }
1505 SHARED_RING_INIT(txs);
1506 FRONT_RING_INIT(&info->tx, txs, PAGE_SIZE);
1507
1508 err = xenbus_grant_ring(dev, virt_to_mfn(txs));
1509 if (err < 0) {
1510 free_page((unsigned long)txs);
1511 goto fail;
1512 }
1513
1514 info->tx_ring_ref = err;
1515 rxs = (struct xen_netif_rx_sring *)get_zeroed_page(GFP_NOIO | __GFP_HIGH);
1516 if (!rxs) {
1517 err = -ENOMEM;
1518 xenbus_dev_fatal(dev, err, "allocating rx ring page");
1519 goto fail;
1520 }
1521 SHARED_RING_INIT(rxs);
1522 FRONT_RING_INIT(&info->rx, rxs, PAGE_SIZE);
1523
1524 err = xenbus_grant_ring(dev, virt_to_mfn(rxs));
1525 if (err < 0) {
1526 free_page((unsigned long)rxs);
1527 goto fail;
1528 }
1529 info->rx_ring_ref = err;
1530
1531 err = xenbus_alloc_evtchn(dev, &info->evtchn);
1532 if (err)
1533 goto fail;
1534
1535 err = bind_evtchn_to_irqhandler(info->evtchn, xennet_interrupt,
1536 0, netdev->name, netdev);
1537 if (err < 0)
1538 goto fail;
1539 netdev->irq = err;
1540 return 0;
1541
1542 fail:
1543 return err;
1544}
1545
1546/* Common code used when first setting up, and when resuming. */
1547static int talk_to_netback(struct xenbus_device *dev,
1548 struct netfront_info *info)
1549{
1550 const char *message;
1551 struct xenbus_transaction xbt;
1552 int err;
1553
1554 /* Create shared ring, alloc event channel. */
1555 err = setup_netfront(dev, info);
1556 if (err)
1557 goto out;
1558
1559again:
1560 err = xenbus_transaction_start(&xbt);
1561 if (err) {
1562 xenbus_dev_fatal(dev, err, "starting transaction");
1563 goto destroy_ring;
1564 }
1565
1566 err = xenbus_printf(xbt, dev->nodename, "tx-ring-ref", "%u",
1567 info->tx_ring_ref);
1568 if (err) {
1569 message = "writing tx ring-ref";
1570 goto abort_transaction;
1571 }
1572 err = xenbus_printf(xbt, dev->nodename, "rx-ring-ref", "%u",
1573 info->rx_ring_ref);
1574 if (err) {
1575 message = "writing rx ring-ref";
1576 goto abort_transaction;
1577 }
1578 err = xenbus_printf(xbt, dev->nodename,
1579 "event-channel", "%u", info->evtchn);
1580 if (err) {
1581 message = "writing event-channel";
1582 goto abort_transaction;
1583 }
1584
1585 err = xenbus_printf(xbt, dev->nodename, "request-rx-copy", "%u",
1586 1);
1587 if (err) {
1588 message = "writing request-rx-copy";
1589 goto abort_transaction;
1590 }
1591
1592 err = xenbus_printf(xbt, dev->nodename, "feature-rx-notify", "%d", 1);
1593 if (err) {
1594 message = "writing feature-rx-notify";
1595 goto abort_transaction;
1596 }
1597
1598 err = xenbus_printf(xbt, dev->nodename, "feature-sg", "%d", 1);
1599 if (err) {
1600 message = "writing feature-sg";
1601 goto abort_transaction;
1602 }
1603
1604 err = xenbus_printf(xbt, dev->nodename, "feature-gso-tcpv4", "%d", 1);
1605 if (err) {
1606 message = "writing feature-gso-tcpv4";
1607 goto abort_transaction;
1608 }
1609
1610 err = xenbus_transaction_end(xbt, 0);
1611 if (err) {
1612 if (err == -EAGAIN)
1613 goto again;
1614 xenbus_dev_fatal(dev, err, "completing transaction");
1615 goto destroy_ring;
1616 }
1617
1618 return 0;
1619
1620 abort_transaction:
1621 xenbus_transaction_end(xbt, 1);
1622 xenbus_dev_fatal(dev, err, "%s", message);
1623 destroy_ring:
1624 xennet_disconnect_backend(info);
1625 out:
1626 return err;
1627}
1628
1629static int xennet_connect(struct net_device *dev)
1630{
1631 struct netfront_info *np = netdev_priv(dev);
1632 int i, requeue_idx, err;
1633 struct sk_buff *skb;
1634 grant_ref_t ref;
1635 struct xen_netif_rx_request *req;
1636 unsigned int feature_rx_copy;
1637
1638 err = xenbus_scanf(XBT_NIL, np->xbdev->otherend,
1639 "feature-rx-copy", "%u", &feature_rx_copy);
1640 if (err != 1)
1641 feature_rx_copy = 0;
1642
1643 if (!feature_rx_copy) {
1644 dev_info(&dev->dev,
1645 "backend does not support copying receive path\n");
1646 return -ENODEV;
1647 }
1648
1649 err = talk_to_netback(np->xbdev, np);
1650 if (err)
1651 return err;
1652
1653 rtnl_lock();
1654 netdev_update_features(dev);
1655 rtnl_unlock();
1656
1657 spin_lock_bh(&np->rx_lock);
1658 spin_lock_irq(&np->tx_lock);
1659
1660 /* Step 1: Discard all pending TX packet fragments. */
1661 xennet_release_tx_bufs(np);
1662
1663 /* Step 2: Rebuild the RX buffer freelist and the RX ring itself. */
1664 for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
1665 skb_frag_t *frag;
1666 const struct page *page;
1667 if (!np->rx_skbs[i])
1668 continue;
1669
1670 skb = np->rx_skbs[requeue_idx] = xennet_get_rx_skb(np, i);
1671 ref = np->grant_rx_ref[requeue_idx] = xennet_get_rx_ref(np, i);
1672 req = RING_GET_REQUEST(&np->rx, requeue_idx);
1673
1674 frag = &skb_shinfo(skb)->frags[0];
1675 page = skb_frag_page(frag);
1676 gnttab_grant_foreign_access_ref(
1677 ref, np->xbdev->otherend_id,
1678 pfn_to_mfn(page_to_pfn(page)),
1679 0);
1680 req->gref = ref;
1681 req->id = requeue_idx;
1682
1683 requeue_idx++;
1684 }
1685
1686 np->rx.req_prod_pvt = requeue_idx;
1687
1688 /*
1689 * Step 3: All public and private state should now be sane. Get
1690 * ready to start sending and receiving packets and give the driver
1691 * domain a kick because we've probably just requeued some
1692 * packets.
1693 */
1694 netif_carrier_on(np->netdev);
1695 notify_remote_via_irq(np->netdev->irq);
1696 xennet_tx_buf_gc(dev);
1697 xennet_alloc_rx_buffers(dev);
1698
1699 spin_unlock_irq(&np->tx_lock);
1700 spin_unlock_bh(&np->rx_lock);
1701
1702 return 0;
1703}
1704
1705/**
1706 * Callback received when the backend's state changes.
1707 */
1708static void netback_changed(struct xenbus_device *dev,
1709 enum xenbus_state backend_state)
1710{
1711 struct netfront_info *np = dev_get_drvdata(&dev->dev);
1712 struct net_device *netdev = np->netdev;
1713
1714 dev_dbg(&dev->dev, "%s\n", xenbus_strstate(backend_state));
1715
1716 switch (backend_state) {
1717 case XenbusStateInitialising:
1718 case XenbusStateInitialised:
1719 case XenbusStateReconfiguring:
1720 case XenbusStateReconfigured:
1721 case XenbusStateUnknown:
1722 case XenbusStateClosed:
1723 break;
1724
1725 case XenbusStateInitWait:
1726 if (dev->state != XenbusStateInitialising)
1727 break;
1728 if (xennet_connect(netdev) != 0)
1729 break;
1730 xenbus_switch_state(dev, XenbusStateConnected);
1731 break;
1732
1733 case XenbusStateConnected:
1734 netif_notify_peers(netdev);
1735 break;
1736
1737 case XenbusStateClosing:
1738 xenbus_frontend_closed(dev);
1739 break;
1740 }
1741}
1742
1743static const struct xennet_stat {
1744 char name[ETH_GSTRING_LEN];
1745 u16 offset;
1746} xennet_stats[] = {
1747 {
1748 "rx_gso_checksum_fixup",
1749 offsetof(struct netfront_info, rx_gso_checksum_fixup)
1750 },
1751};
1752
1753static int xennet_get_sset_count(struct net_device *dev, int string_set)
1754{
1755 switch (string_set) {
1756 case ETH_SS_STATS:
1757 return ARRAY_SIZE(xennet_stats);
1758 default:
1759 return -EINVAL;
1760 }
1761}
1762
1763static void xennet_get_ethtool_stats(struct net_device *dev,
1764 struct ethtool_stats *stats, u64 * data)
1765{
1766 void *np = netdev_priv(dev);
1767 int i;
1768
1769 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++)
1770 data[i] = *(unsigned long *)(np + xennet_stats[i].offset);
1771}
1772
1773static void xennet_get_strings(struct net_device *dev, u32 stringset, u8 * data)
1774{
1775 int i;
1776
1777 switch (stringset) {
1778 case ETH_SS_STATS:
1779 for (i = 0; i < ARRAY_SIZE(xennet_stats); i++)
1780 memcpy(data + i * ETH_GSTRING_LEN,
1781 xennet_stats[i].name, ETH_GSTRING_LEN);
1782 break;
1783 }
1784}
1785
1786static const struct ethtool_ops xennet_ethtool_ops =
1787{
1788 .get_link = ethtool_op_get_link,
1789
1790 .get_sset_count = xennet_get_sset_count,
1791 .get_ethtool_stats = xennet_get_ethtool_stats,
1792 .get_strings = xennet_get_strings,
1793};
1794
1795#ifdef CONFIG_SYSFS
1796static ssize_t show_rxbuf_min(struct device *dev,
1797 struct device_attribute *attr, char *buf)
1798{
1799 struct net_device *netdev = to_net_dev(dev);
1800 struct netfront_info *info = netdev_priv(netdev);
1801
1802 return sprintf(buf, "%u\n", info->rx_min_target);
1803}
1804
1805static ssize_t store_rxbuf_min(struct device *dev,
1806 struct device_attribute *attr,
1807 const char *buf, size_t len)
1808{
1809 struct net_device *netdev = to_net_dev(dev);
1810 struct netfront_info *np = netdev_priv(netdev);
1811 char *endp;
1812 unsigned long target;
1813
1814 if (!capable(CAP_NET_ADMIN))
1815 return -EPERM;
1816
1817 target = simple_strtoul(buf, &endp, 0);
1818 if (endp == buf)
1819 return -EBADMSG;
1820
1821 if (target < RX_MIN_TARGET)
1822 target = RX_MIN_TARGET;
1823 if (target > RX_MAX_TARGET)
1824 target = RX_MAX_TARGET;
1825
1826 spin_lock_bh(&np->rx_lock);
1827 if (target > np->rx_max_target)
1828 np->rx_max_target = target;
1829 np->rx_min_target = target;
1830 if (target > np->rx_target)
1831 np->rx_target = target;
1832
1833 xennet_alloc_rx_buffers(netdev);
1834
1835 spin_unlock_bh(&np->rx_lock);
1836 return len;
1837}
1838
1839static ssize_t show_rxbuf_max(struct device *dev,
1840 struct device_attribute *attr, char *buf)
1841{
1842 struct net_device *netdev = to_net_dev(dev);
1843 struct netfront_info *info = netdev_priv(netdev);
1844
1845 return sprintf(buf, "%u\n", info->rx_max_target);
1846}
1847
1848static ssize_t store_rxbuf_max(struct device *dev,
1849 struct device_attribute *attr,
1850 const char *buf, size_t len)
1851{
1852 struct net_device *netdev = to_net_dev(dev);
1853 struct netfront_info *np = netdev_priv(netdev);
1854 char *endp;
1855 unsigned long target;
1856
1857 if (!capable(CAP_NET_ADMIN))
1858 return -EPERM;
1859
1860 target = simple_strtoul(buf, &endp, 0);
1861 if (endp == buf)
1862 return -EBADMSG;
1863
1864 if (target < RX_MIN_TARGET)
1865 target = RX_MIN_TARGET;
1866 if (target > RX_MAX_TARGET)
1867 target = RX_MAX_TARGET;
1868
1869 spin_lock_bh(&np->rx_lock);
1870 if (target < np->rx_min_target)
1871 np->rx_min_target = target;
1872 np->rx_max_target = target;
1873 if (target < np->rx_target)
1874 np->rx_target = target;
1875
1876 xennet_alloc_rx_buffers(netdev);
1877
1878 spin_unlock_bh(&np->rx_lock);
1879 return len;
1880}
1881
1882static ssize_t show_rxbuf_cur(struct device *dev,
1883 struct device_attribute *attr, char *buf)
1884{
1885 struct net_device *netdev = to_net_dev(dev);
1886 struct netfront_info *info = netdev_priv(netdev);
1887
1888 return sprintf(buf, "%u\n", info->rx_target);
1889}
1890
1891static struct device_attribute xennet_attrs[] = {
1892 __ATTR(rxbuf_min, S_IRUGO|S_IWUSR, show_rxbuf_min, store_rxbuf_min),
1893 __ATTR(rxbuf_max, S_IRUGO|S_IWUSR, show_rxbuf_max, store_rxbuf_max),
1894 __ATTR(rxbuf_cur, S_IRUGO, show_rxbuf_cur, NULL),
1895};
1896
1897static int xennet_sysfs_addif(struct net_device *netdev)
1898{
1899 int i;
1900 int err;
1901
1902 for (i = 0; i < ARRAY_SIZE(xennet_attrs); i++) {
1903 err = device_create_file(&netdev->dev,
1904 &xennet_attrs[i]);
1905 if (err)
1906 goto fail;
1907 }
1908 return 0;
1909
1910 fail:
1911 while (--i >= 0)
1912 device_remove_file(&netdev->dev, &xennet_attrs[i]);
1913 return err;
1914}
1915
1916static void xennet_sysfs_delif(struct net_device *netdev)
1917{
1918 int i;
1919
1920 for (i = 0; i < ARRAY_SIZE(xennet_attrs); i++)
1921 device_remove_file(&netdev->dev, &xennet_attrs[i]);
1922}
1923
1924#endif /* CONFIG_SYSFS */
1925
1926static const struct xenbus_device_id netfront_ids[] = {
1927 { "vif" },
1928 { "" }
1929};
1930
1931
1932static int __devexit xennet_remove(struct xenbus_device *dev)
1933{
1934 struct netfront_info *info = dev_get_drvdata(&dev->dev);
1935
1936 dev_dbg(&dev->dev, "%s\n", dev->nodename);
1937
1938 xennet_disconnect_backend(info);
1939
1940 xennet_sysfs_delif(info->netdev);
1941
1942 unregister_netdev(info->netdev);
1943
1944 del_timer_sync(&info->rx_refill_timer);
1945
1946 free_percpu(info->stats);
1947
1948 free_netdev(info->netdev);
1949
1950 return 0;
1951}
1952
1953static DEFINE_XENBUS_DRIVER(netfront, ,
1954 .probe = netfront_probe,
1955 .remove = __devexit_p(xennet_remove),
1956 .resume = netfront_resume,
1957 .otherend_changed = netback_changed,
1958);
1959
1960static int __init netif_init(void)
1961{
1962 if (!xen_domain())
1963 return -ENODEV;
1964
1965 if (xen_hvm_domain() && !xen_platform_pci_unplug)
1966 return -ENODEV;
1967
1968 printk(KERN_INFO "Initialising Xen virtual ethernet driver.\n");
1969
1970 return xenbus_register_frontend(&netfront_driver);
1971}
1972module_init(netif_init);
1973
1974
1975static void __exit netif_exit(void)
1976{
1977 xenbus_unregister_driver(&netfront_driver);
1978}
1979module_exit(netif_exit);
1980
1981MODULE_DESCRIPTION("Xen virtual network device frontend");
1982MODULE_LICENSE("GPL");
1983MODULE_ALIAS("xen:vif");
1984MODULE_ALIAS("xennet");