?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge remote-tracking branch 'cleancache/linux-next'
[linux-2.6/next.git] / drivers / net / xen-netback / netback.c
blob0e4851b8a773642322854be32fe764747980098e
1 /*
2 * Back-end of the driver for virtual network devices. This portion of the
3 * driver exports a 'unified' network-device interface that can be accessed
4 * by any operating system that implements a compatible front end. A
5 * reference front-end implementation can be found in:
6 * drivers/net/xen-netfront.c
7 *
8 * Copyright (c) 2002-2005, K A Fraser
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public License version 2
12 * as published by the Free Software Foundation; or, when distributed
13 * separately from the Linux kernel or incorporated into other
14 * software packages, subject to the following license:
15 *
16 * Permission is hereby granted, free of charge, to any person obtaining a copy
17 * of this source file (the "Software"), to deal in the Software without
18 * restriction, including without limitation the rights to use, copy, modify,
19 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
20 * and to permit persons to whom the Software is furnished to do so, subject to
21 * the following conditions:
22 *
23 * The above copyright notice and this permission notice shall be included in
24 * all copies or substantial portions of the Software.
25 *
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
27 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
28 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
29 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
30 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
31 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
32 * IN THE SOFTWARE.
33 */
34
35 #include "common.h"
36
37 #include <linux/kthread.h>
38 #include <linux/if_vlan.h>
39 #include <linux/udp.h>
40
41 #include <net/tcp.h>
42
43 #include <xen/events.h>
44 #include <xen/interface/memory.h>
45
46 #include <asm/xen/hypercall.h>
47 #include <asm/xen/page.h>
48
49 struct pending_tx_info {
50 struct xen_netif_tx_request req;
51 struct xenvif *vif;
52 };
53 typedef unsigned int pending_ring_idx_t;
54
55 struct netbk_rx_meta {
56 int id;
57 int size;
58 int gso_size;
59 };
60
61 #define MAX_PENDING_REQS 256
62
63 #define MAX_BUFFER_OFFSET PAGE_SIZE
64
65 /* extra field used in struct page */
66 union page_ext {
67 struct {
68 #if BITS_PER_LONG < 64
69 #define IDX_WIDTH 8
70 #define GROUP_WIDTH (BITS_PER_LONG - IDX_WIDTH)
71 unsigned int group:GROUP_WIDTH;
72 unsigned int idx:IDX_WIDTH;
73 #else
74 unsigned int group, idx;
75 #endif
76 } e;
77 void *mapping;
78 };
79
80 struct xen_netbk {
81 wait_queue_head_t wq;
82 struct task_struct *task;
83
84 struct sk_buff_head rx_queue;
85 struct sk_buff_head tx_queue;
86
87 struct timer_list net_timer;
88
89 struct page *mmap_pages[MAX_PENDING_REQS];
90
91 pending_ring_idx_t pending_prod;
92 pending_ring_idx_t pending_cons;
93 struct list_head net_schedule_list;
94
95 /* Protect the net_schedule_list in netif. */
96 spinlock_t net_schedule_list_lock;
97
98 atomic_t netfront_count;
99
100 struct pending_tx_info pending_tx_info[MAX_PENDING_REQS];
101 struct gnttab_copy tx_copy_ops[MAX_PENDING_REQS];
102
103 u16 pending_ring[MAX_PENDING_REQS];
104
105 /*
106 * Given MAX_BUFFER_OFFSET of 4096 the worst case is that each
107 * head/fragment page uses 2 copy operations because it
108 * straddles two buffers in the frontend.
109 */
110 struct gnttab_copy grant_copy_op[2*XEN_NETIF_RX_RING_SIZE];
111 struct netbk_rx_meta meta[2*XEN_NETIF_RX_RING_SIZE];
112 };
113
114 static struct xen_netbk *xen_netbk;
115 static int xen_netbk_group_nr;
116
117 void xen_netbk_add_xenvif(struct xenvif *vif)
118 {
119 int i;
120 int min_netfront_count;
121 int min_group = 0;
122 struct xen_netbk *netbk;
123
124 min_netfront_count = atomic_read(&xen_netbk[0].netfront_count);
125 for (i = 0; i < xen_netbk_group_nr; i++) {
126 int netfront_count = atomic_read(&xen_netbk[i].netfront_count);
127 if (netfront_count < min_netfront_count) {
128 min_group = i;
129 min_netfront_count = netfront_count;
130 }
131 }
132
133 netbk = &xen_netbk[min_group];
134
135 vif->netbk = netbk;
136 atomic_inc(&netbk->netfront_count);
137 }
138
139 void xen_netbk_remove_xenvif(struct xenvif *vif)
140 {
141 struct xen_netbk *netbk = vif->netbk;
142 vif->netbk = NULL;
143 atomic_dec(&netbk->netfront_count);
144 }
145
146 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx);
147 static void make_tx_response(struct xenvif *vif,
148 struct xen_netif_tx_request *txp,
149 s8 st);
150 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
151 u16 id,
152 s8 st,
153 u16 offset,
154 u16 size,
155 u16 flags);
156
157 static inline unsigned long idx_to_pfn(struct xen_netbk *netbk,
158 unsigned int idx)
159 {
160 return page_to_pfn(netbk->mmap_pages[idx]);
161 }
162
163 static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
164 unsigned int idx)
165 {
166 return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
167 }
168
169 /* extra field used in struct page */
170 static inline void set_page_ext(struct page *pg, struct xen_netbk *netbk,
171 unsigned int idx)
172 {
173 unsigned int group = netbk - xen_netbk;
174 union page_ext ext = { .e = { .group = group + 1, .idx = idx } };
175
176 BUILD_BUG_ON(sizeof(ext) > sizeof(ext.mapping));
177 pg->mapping = ext.mapping;
178 }
179
180 static int get_page_ext(struct page *pg,
181 unsigned int *pgroup, unsigned int *pidx)
182 {
183 union page_ext ext = { .mapping = pg->mapping };
184 struct xen_netbk *netbk;
185 unsigned int group, idx;
186
187 group = ext.e.group - 1;
188
189 if (group < 0 || group >= xen_netbk_group_nr)
190 return 0;
191
192 netbk = &xen_netbk[group];
193
194 idx = ext.e.idx;
195
196 if ((idx < 0) || (idx >= MAX_PENDING_REQS))
197 return 0;
198
199 if (netbk->mmap_pages[idx] != pg)
200 return 0;
201
202 *pgroup = group;
203 *pidx = idx;
204
205 return 1;
206 }
207
208 /*
209 * This is the amount of packet we copy rather than map, so that the
210 * guest can't fiddle with the contents of the headers while we do
211 * packet processing on them (netfilter, routing, etc).
212 */
213 #define PKT_PROT_LEN (ETH_HLEN + \
214 VLAN_HLEN + \
215 sizeof(struct iphdr) + MAX_IPOPTLEN + \
216 sizeof(struct tcphdr) + MAX_TCP_OPTION_SPACE)
217
218 static inline pending_ring_idx_t pending_index(unsigned i)
219 {
220 return i & (MAX_PENDING_REQS-1);
221 }
222
223 static inline pending_ring_idx_t nr_pending_reqs(struct xen_netbk *netbk)
224 {
225 return MAX_PENDING_REQS -
226 netbk->pending_prod + netbk->pending_cons;
227 }
228
229 static void xen_netbk_kick_thread(struct xen_netbk *netbk)
230 {
231 wake_up(&netbk->wq);
232 }
233
234 static int max_required_rx_slots(struct xenvif *vif)
235 {
236 int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
237
238 if (vif->can_sg || vif->gso || vif->gso_prefix)
239 max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
240
241 return max;
242 }
243
244 int xen_netbk_rx_ring_full(struct xenvif *vif)
245 {
246 RING_IDX peek = vif->rx_req_cons_peek;
247 RING_IDX needed = max_required_rx_slots(vif);
248
249 return ((vif->rx.sring->req_prod - peek) < needed) ||
250 ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
251 }
252
253 int xen_netbk_must_stop_queue(struct xenvif *vif)
254 {
255 if (!xen_netbk_rx_ring_full(vif))
256 return 0;
257
258 vif->rx.sring->req_event = vif->rx_req_cons_peek +
259 max_required_rx_slots(vif);
260 mb(); /* request notification /then/ check the queue */
261
262 return xen_netbk_rx_ring_full(vif);
263 }
264
265 /*
266 * Returns true if we should start a new receive buffer instead of
267 * adding 'size' bytes to a buffer which currently contains 'offset'
268 * bytes.
269 */
270 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
271 {
272 /* simple case: we have completely filled the current buffer. */
273 if (offset == MAX_BUFFER_OFFSET)
274 return true;
275
276 /*
277 * complex case: start a fresh buffer if the current frag
278 * would overflow the current buffer but only if:
279 * (i) this frag would fit completely in the next buffer
280 * and (ii) there is already some data in the current buffer
281 * and (iii) this is not the head buffer.
282 *
283 * Where:
284 * - (i) stops us splitting a frag into two copies
285 * unless the frag is too large for a single buffer.
286 * - (ii) stops us from leaving a buffer pointlessly empty.
287 * - (iii) stops us leaving the first buffer
288 * empty. Strictly speaking this is already covered
289 * by (ii) but is explicitly checked because
290 * netfront relies on the first buffer being
291 * non-empty and can crash otherwise.
292 *
293 * This means we will effectively linearise small
294 * frags but do not needlessly split large buffers
295 * into multiple copies tend to give large frags their
296 * own buffers as before.
297 */
298 if ((offset + size > MAX_BUFFER_OFFSET) &&
299 (size <= MAX_BUFFER_OFFSET) && offset && !head)
300 return true;
301
302 return false;
303 }
304
305 /*
306 * Figure out how many ring slots we're going to need to send @skb to
307 * the guest. This function is essentially a dry run of
308 * netbk_gop_frag_copy.
309 */
310 unsigned int xen_netbk_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
311 {
312 unsigned int count;
313 int i, copy_off;
314
315 count = DIV_ROUND_UP(
316 offset_in_page(skb->data)+skb_headlen(skb), PAGE_SIZE);
317
318 copy_off = skb_headlen(skb) % PAGE_SIZE;
319
320 if (skb_shinfo(skb)->gso_size)
321 count++;
322
323 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
324 unsigned long size = skb_shinfo(skb)->frags[i].size;
325 unsigned long bytes;
326 while (size > 0) {
327 BUG_ON(copy_off > MAX_BUFFER_OFFSET);
328
329 if (start_new_rx_buffer(copy_off, size, 0)) {
330 count++;
331 copy_off = 0;
332 }
333
334 bytes = size;
335 if (copy_off + bytes > MAX_BUFFER_OFFSET)
336 bytes = MAX_BUFFER_OFFSET - copy_off;
337
338 copy_off += bytes;
339 size -= bytes;
340 }
341 }
342 return count;
343 }
344
345 struct netrx_pending_operations {
346 unsigned copy_prod, copy_cons;
347 unsigned meta_prod, meta_cons;
348 struct gnttab_copy *copy;
349 struct netbk_rx_meta *meta;
350 int copy_off;
351 grant_ref_t copy_gref;
352 };
353
354 static struct netbk_rx_meta *get_next_rx_buffer(struct xenvif *vif,
355 struct netrx_pending_operations *npo)
356 {
357 struct netbk_rx_meta *meta;
358 struct xen_netif_rx_request *req;
359
360 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
361
362 meta = npo->meta + npo->meta_prod++;
363 meta->gso_size = 0;
364 meta->size = 0;
365 meta->id = req->id;
366
367 npo->copy_off = 0;
368 npo->copy_gref = req->gref;
369
370 return meta;
371 }
372
373 /*
374 * Set up the grant operations for this fragment. If it's a flipping
375 * interface, we also set up the unmap request from here.
376 */
377 static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
378 struct netrx_pending_operations *npo,
379 struct page *page, unsigned long size,
380 unsigned long offset, int *head)
381 {
382 struct gnttab_copy *copy_gop;
383 struct netbk_rx_meta *meta;
384 /*
385 * These variables a used iff get_page_ext returns true,
386 * in which case they are guaranteed to be initialized.
387 */
388 unsigned int uninitialized_var(group), uninitialized_var(idx);
389 int foreign = get_page_ext(page, &group, &idx);
390 unsigned long bytes;
391
392 /* Data must not cross a page boundary. */
393 BUG_ON(size + offset > PAGE_SIZE);
394
395 meta = npo->meta + npo->meta_prod - 1;
396
397 while (size > 0) {
398 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
399
400 if (start_new_rx_buffer(npo->copy_off, size, *head)) {
401 /*
402 * Netfront requires there to be some data in the head
403 * buffer.
404 */
405 BUG_ON(*head);
406
407 meta = get_next_rx_buffer(vif, npo);
408 }
409
410 bytes = size;
411 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
412 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
413
414 copy_gop = npo->copy + npo->copy_prod++;
415 copy_gop->flags = GNTCOPY_dest_gref;
416 if (foreign) {
417 struct xen_netbk *netbk = &xen_netbk[group];
418 struct pending_tx_info *src_pend;
419
420 src_pend = &netbk->pending_tx_info[idx];
421
422 copy_gop->source.domid = src_pend->vif->domid;
423 copy_gop->source.u.ref = src_pend->req.gref;
424 copy_gop->flags |= GNTCOPY_source_gref;
425 } else {
426 void *vaddr = page_address(page);
427 copy_gop->source.domid = DOMID_SELF;
428 copy_gop->source.u.gmfn = virt_to_mfn(vaddr);
429 }
430 copy_gop->source.offset = offset;
431 copy_gop->dest.domid = vif->domid;
432
433 copy_gop->dest.offset = npo->copy_off;
434 copy_gop->dest.u.ref = npo->copy_gref;
435 copy_gop->len = bytes;
436
437 npo->copy_off += bytes;
438 meta->size += bytes;
439
440 offset += bytes;
441 size -= bytes;
442
443 /* Leave a gap for the GSO descriptor. */
444 if (*head && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
445 vif->rx.req_cons++;
446
447 *head = 0; /* There must be something in this buffer now. */
448
449 }
450 }
451
452 /*
453 * Prepare an SKB to be transmitted to the frontend.
454 *
455 * This function is responsible for allocating grant operations, meta
456 * structures, etc.
457 *
458 * It returns the number of meta structures consumed. The number of
459 * ring slots used is always equal to the number of meta slots used
460 * plus the number of GSO descriptors used. Currently, we use either
461 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
462 * frontend-side LRO).
463 */
464 static int netbk_gop_skb(struct sk_buff *skb,
465 struct netrx_pending_operations *npo)
466 {
467 struct xenvif *vif = netdev_priv(skb->dev);
468 int nr_frags = skb_shinfo(skb)->nr_frags;
469 int i;
470 struct xen_netif_rx_request *req;
471 struct netbk_rx_meta *meta;
472 unsigned char *data;
473 int head = 1;
474 int old_meta_prod;
475
476 old_meta_prod = npo->meta_prod;
477
478 /* Set up a GSO prefix descriptor, if necessary */
479 if (skb_shinfo(skb)->gso_size && vif->gso_prefix) {
480 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
481 meta = npo->meta + npo->meta_prod++;
482 meta->gso_size = skb_shinfo(skb)->gso_size;
483 meta->size = 0;
484 meta->id = req->id;
485 }
486
487 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
488 meta = npo->meta + npo->meta_prod++;
489
490 if (!vif->gso_prefix)
491 meta->gso_size = skb_shinfo(skb)->gso_size;
492 else
493 meta->gso_size = 0;
494
495 meta->size = 0;
496 meta->id = req->id;
497 npo->copy_off = 0;
498 npo->copy_gref = req->gref;
499
500 data = skb->data;
501 while (data < skb_tail_pointer(skb)) {
502 unsigned int offset = offset_in_page(data);
503 unsigned int len = PAGE_SIZE - offset;
504
505 if (data + len > skb_tail_pointer(skb))
506 len = skb_tail_pointer(skb) - data;
507
508 netbk_gop_frag_copy(vif, skb, npo,
509 virt_to_page(data), len, offset, &head);
510 data += len;
511 }
512
513 for (i = 0; i < nr_frags; i++) {
514 netbk_gop_frag_copy(vif, skb, npo,
515 skb_shinfo(skb)->frags[i].page,
516 skb_shinfo(skb)->frags[i].size,
517 skb_shinfo(skb)->frags[i].page_offset,
518 &head);
519 }
520
521 return npo->meta_prod - old_meta_prod;
522 }
523
524 /*
525 * This is a twin to netbk_gop_skb. Assume that netbk_gop_skb was
526 * used to set up the operations on the top of
527 * netrx_pending_operations, which have since been done. Check that
528 * they didn't give any errors and advance over them.
529 */
530 static int netbk_check_gop(struct xenvif *vif, int nr_meta_slots,
531 struct netrx_pending_operations *npo)
532 {
533 struct gnttab_copy *copy_op;
534 int status = XEN_NETIF_RSP_OKAY;
535 int i;
536
537 for (i = 0; i < nr_meta_slots; i++) {
538 copy_op = npo->copy + npo->copy_cons++;
539 if (copy_op->status != GNTST_okay) {
540 netdev_dbg(vif->dev,
541 "Bad status %d from copy to DOM%d.\n",
542 copy_op->status, vif->domid);
543 status = XEN_NETIF_RSP_ERROR;
544 }
545 }
546
547 return status;
548 }
549
550 static void netbk_add_frag_responses(struct xenvif *vif, int status,
551 struct netbk_rx_meta *meta,
552 int nr_meta_slots)
553 {
554 int i;
555 unsigned long offset;
556
557 /* No fragments used */
558 if (nr_meta_slots <= 1)
559 return;
560
561 nr_meta_slots--;
562
563 for (i = 0; i < nr_meta_slots; i++) {
564 int flags;
565 if (i == nr_meta_slots - 1)
566 flags = 0;
567 else
568 flags = XEN_NETRXF_more_data;
569
570 offset = 0;
571 make_rx_response(vif, meta[i].id, status, offset,
572 meta[i].size, flags);
573 }
574 }
575
576 struct skb_cb_overlay {
577 int meta_slots_used;
578 };
579
580 static void xen_netbk_rx_action(struct xen_netbk *netbk)
581 {
582 struct xenvif *vif = NULL, *tmp;
583 s8 status;
584 u16 irq, flags;
585 struct xen_netif_rx_response *resp;
586 struct sk_buff_head rxq;
587 struct sk_buff *skb;
588 LIST_HEAD(notify);
589 int ret;
590 int nr_frags;
591 int count;
592 unsigned long offset;
593 struct skb_cb_overlay *sco;
594
595 struct netrx_pending_operations npo = {
596 .copy = netbk->grant_copy_op,
597 .meta = netbk->meta,
598 };
599
600 skb_queue_head_init(&rxq);
601
602 count = 0;
603
604 while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
605 vif = netdev_priv(skb->dev);
606 nr_frags = skb_shinfo(skb)->nr_frags;
607
608 sco = (struct skb_cb_overlay *)skb->cb;
609 sco->meta_slots_used = netbk_gop_skb(skb, &npo);
610
611 count += nr_frags + 1;
612
613 __skb_queue_tail(&rxq, skb);
614
615 /* Filled the batch queue? */
616 if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
617 break;
618 }
619
620 BUG_ON(npo.meta_prod > ARRAY_SIZE(netbk->meta));
621
622 if (!npo.copy_prod)
623 return;
624
625 BUG_ON(npo.copy_prod > ARRAY_SIZE(netbk->grant_copy_op));
626 ret = HYPERVISOR_grant_table_op(GNTTABOP_copy, &netbk->grant_copy_op,
627 npo.copy_prod);
628 BUG_ON(ret != 0);
629
630 while ((skb = __skb_dequeue(&rxq)) != NULL) {
631 sco = (struct skb_cb_overlay *)skb->cb;
632
633 vif = netdev_priv(skb->dev);
634
635 if (netbk->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
636 resp = RING_GET_RESPONSE(&vif->rx,
637 vif->rx.rsp_prod_pvt++);
638
639 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
640
641 resp->offset = netbk->meta[npo.meta_cons].gso_size;
642 resp->id = netbk->meta[npo.meta_cons].id;
643 resp->status = sco->meta_slots_used;
644
645 npo.meta_cons++;
646 sco->meta_slots_used--;
647 }
648
649
650 vif->dev->stats.tx_bytes += skb->len;
651 vif->dev->stats.tx_packets++;
652
653 status = netbk_check_gop(vif, sco->meta_slots_used, &npo);
654
655 if (sco->meta_slots_used == 1)
656 flags = 0;
657 else
658 flags = XEN_NETRXF_more_data;
659
660 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
661 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
662 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
663 /* remote but checksummed. */
664 flags |= XEN_NETRXF_data_validated;
665
666 offset = 0;
667 resp = make_rx_response(vif, netbk->meta[npo.meta_cons].id,
668 status, offset,
669 netbk->meta[npo.meta_cons].size,
670 flags);
671
672 if (netbk->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
673 struct xen_netif_extra_info *gso =
674 (struct xen_netif_extra_info *)
675 RING_GET_RESPONSE(&vif->rx,
676 vif->rx.rsp_prod_pvt++);
677
678 resp->flags |= XEN_NETRXF_extra_info;
679
680 gso->u.gso.size = netbk->meta[npo.meta_cons].gso_size;
681 gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
682 gso->u.gso.pad = 0;
683 gso->u.gso.features = 0;
684
685 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
686 gso->flags = 0;
687 }
688
689 netbk_add_frag_responses(vif, status,
690 netbk->meta + npo.meta_cons + 1,
691 sco->meta_slots_used);
692
693 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
694 irq = vif->irq;
695 if (ret && list_empty(&vif->notify_list))
696 list_add_tail(&vif->notify_list, &notify);
697
698 xenvif_notify_tx_completion(vif);
699
700 xenvif_put(vif);
701 npo.meta_cons += sco->meta_slots_used;
702 dev_kfree_skb(skb);
703 }
704
705 list_for_each_entry_safe(vif, tmp, &notify, notify_list) {
706 notify_remote_via_irq(vif->irq);
707 list_del_init(&vif->notify_list);
708 }
709
710 /* More work to do? */
711 if (!skb_queue_empty(&netbk->rx_queue) &&
712 !timer_pending(&netbk->net_timer))
713 xen_netbk_kick_thread(netbk);
714 }
715
716 void xen_netbk_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
717 {
718 struct xen_netbk *netbk = vif->netbk;
719
720 skb_queue_tail(&netbk->rx_queue, skb);
721
722 xen_netbk_kick_thread(netbk);
723 }
724
725 static void xen_netbk_alarm(unsigned long data)
726 {
727 struct xen_netbk *netbk = (struct xen_netbk *)data;
728 xen_netbk_kick_thread(netbk);
729 }
730
731 static int __on_net_schedule_list(struct xenvif *vif)
732 {
733 return !list_empty(&vif->schedule_list);
734 }
735
736 /* Must be called with net_schedule_list_lock held */
737 static void remove_from_net_schedule_list(struct xenvif *vif)
738 {
739 if (likely(__on_net_schedule_list(vif))) {
740 list_del_init(&vif->schedule_list);
741 xenvif_put(vif);
742 }
743 }
744
745 static struct xenvif *poll_net_schedule_list(struct xen_netbk *netbk)
746 {
747 struct xenvif *vif = NULL;
748
749 spin_lock_irq(&netbk->net_schedule_list_lock);
750 if (list_empty(&netbk->net_schedule_list))
751 goto out;
752
753 vif = list_first_entry(&netbk->net_schedule_list,
754 struct xenvif, schedule_list);
755 if (!vif)
756 goto out;
757
758 xenvif_get(vif);
759
760 remove_from_net_schedule_list(vif);
761 out:
762 spin_unlock_irq(&netbk->net_schedule_list_lock);
763 return vif;
764 }
765
766 void xen_netbk_schedule_xenvif(struct xenvif *vif)
767 {
768 unsigned long flags;
769 struct xen_netbk *netbk = vif->netbk;
770
771 if (__on_net_schedule_list(vif))
772 goto kick;
773
774 spin_lock_irqsave(&netbk->net_schedule_list_lock, flags);
775 if (!__on_net_schedule_list(vif) &&
776 likely(xenvif_schedulable(vif))) {
777 list_add_tail(&vif->schedule_list, &netbk->net_schedule_list);
778 xenvif_get(vif);
779 }
780 spin_unlock_irqrestore(&netbk->net_schedule_list_lock, flags);
781
782 kick:
783 smp_mb();
784 if ((nr_pending_reqs(netbk) < (MAX_PENDING_REQS/2)) &&
785 !list_empty(&netbk->net_schedule_list))
786 xen_netbk_kick_thread(netbk);
787 }
788
789 void xen_netbk_deschedule_xenvif(struct xenvif *vif)
790 {
791 struct xen_netbk *netbk = vif->netbk;
792 spin_lock_irq(&netbk->net_schedule_list_lock);
793 remove_from_net_schedule_list(vif);
794 spin_unlock_irq(&netbk->net_schedule_list_lock);
795 }
796
797 void xen_netbk_check_rx_xenvif(struct xenvif *vif)
798 {
799 int more_to_do;
800
801 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
802
803 if (more_to_do)
804 xen_netbk_schedule_xenvif(vif);
805 }
806
807 static void tx_add_credit(struct xenvif *vif)
808 {
809 unsigned long max_burst, max_credit;
810
811 /*
812 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
813 * Otherwise the interface can seize up due to insufficient credit.
814 */
815 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
816 max_burst = min(max_burst, 131072UL);
817 max_burst = max(max_burst, vif->credit_bytes);
818
819 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
820 max_credit = vif->remaining_credit + vif->credit_bytes;
821 if (max_credit < vif->remaining_credit)
822 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
823
824 vif->remaining_credit = min(max_credit, max_burst);
825 }
826
827 static void tx_credit_callback(unsigned long data)
828 {
829 struct xenvif *vif = (struct xenvif *)data;
830 tx_add_credit(vif);
831 xen_netbk_check_rx_xenvif(vif);
832 }
833
834 static void netbk_tx_err(struct xenvif *vif,
835 struct xen_netif_tx_request *txp, RING_IDX end)
836 {
837 RING_IDX cons = vif->tx.req_cons;
838
839 do {
840 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
841 if (cons >= end)
842 break;
843 txp = RING_GET_REQUEST(&vif->tx, cons++);
844 } while (1);
845 vif->tx.req_cons = cons;
846 xen_netbk_check_rx_xenvif(vif);
847 xenvif_put(vif);
848 }
849
850 static int netbk_count_requests(struct xenvif *vif,
851 struct xen_netif_tx_request *first,
852 struct xen_netif_tx_request *txp,
853 int work_to_do)
854 {
855 RING_IDX cons = vif->tx.req_cons;
856 int frags = 0;
857
858 if (!(first->flags & XEN_NETTXF_more_data))
859 return 0;
860
861 do {
862 if (frags >= work_to_do) {
863 netdev_dbg(vif->dev, "Need more frags\n");
864 return -frags;
865 }
866
867 if (unlikely(frags >= MAX_SKB_FRAGS)) {
868 netdev_dbg(vif->dev, "Too many frags\n");
869 return -frags;
870 }
871
872 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + frags),
873 sizeof(*txp));
874 if (txp->size > first->size) {
875 netdev_dbg(vif->dev, "Frags galore\n");
876 return -frags;
877 }
878
879 first->size -= txp->size;
880 frags++;
881
882 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
883 netdev_dbg(vif->dev, "txp->offset: %x, size: %u\n",
884 txp->offset, txp->size);
885 return -frags;
886 }
887 } while ((txp++)->flags & XEN_NETTXF_more_data);
888 return frags;
889 }
890
891 static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
892 struct sk_buff *skb,
893 unsigned long pending_idx)
894 {
895 struct page *page;
896 page = alloc_page(GFP_KERNEL|__GFP_COLD);
897 if (!page)
898 return NULL;
899 set_page_ext(page, netbk, pending_idx);
900 netbk->mmap_pages[pending_idx] = page;
901 return page;
902 }
903
904 static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
905 struct xenvif *vif,
906 struct sk_buff *skb,
907 struct xen_netif_tx_request *txp,
908 struct gnttab_copy *gop)
909 {
910 struct skb_shared_info *shinfo = skb_shinfo(skb);
911 skb_frag_t *frags = shinfo->frags;
912 unsigned long pending_idx = *((u16 *)skb->data);
913 int i, start;
914
915 /* Skip first skb fragment if it is on same page as header fragment. */
916 start = ((unsigned long)shinfo->frags[0].page == pending_idx);
917
918 for (i = start; i < shinfo->nr_frags; i++, txp++) {
919 struct page *page;
920 pending_ring_idx_t index;
921 struct pending_tx_info *pending_tx_info =
922 netbk->pending_tx_info;
923
924 index = pending_index(netbk->pending_cons++);
925 pending_idx = netbk->pending_ring[index];
926 page = xen_netbk_alloc_page(netbk, skb, pending_idx);
927 if (!page)
928 return NULL;
929
930 netbk->mmap_pages[pending_idx] = page;
931
932 gop->source.u.ref = txp->gref;
933 gop->source.domid = vif->domid;
934 gop->source.offset = txp->offset;
935
936 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
937 gop->dest.domid = DOMID_SELF;
938 gop->dest.offset = txp->offset;
939
940 gop->len = txp->size;
941 gop->flags = GNTCOPY_source_gref;
942
943 gop++;
944
945 memcpy(&pending_tx_info[pending_idx].req, txp, sizeof(*txp));
946 xenvif_get(vif);
947 pending_tx_info[pending_idx].vif = vif;
948 frags[i].page = (void *)pending_idx;
949 }
950
951 return gop;
952 }
953
954 static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
955 struct sk_buff *skb,
956 struct gnttab_copy **gopp)
957 {
958 struct gnttab_copy *gop = *gopp;
959 int pending_idx = *((u16 *)skb->data);
960 struct pending_tx_info *pending_tx_info = netbk->pending_tx_info;
961 struct xenvif *vif = pending_tx_info[pending_idx].vif;
962 struct xen_netif_tx_request *txp;
963 struct skb_shared_info *shinfo = skb_shinfo(skb);
964 int nr_frags = shinfo->nr_frags;
965 int i, err, start;
966
967 /* Check status of header. */
968 err = gop->status;
969 if (unlikely(err)) {
970 pending_ring_idx_t index;
971 index = pending_index(netbk->pending_prod++);
972 txp = &pending_tx_info[pending_idx].req;
973 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
974 netbk->pending_ring[index] = pending_idx;
975 xenvif_put(vif);
976 }
977
978 /* Skip first skb fragment if it is on same page as header fragment. */
979 start = ((unsigned long)shinfo->frags[0].page == pending_idx);
980
981 for (i = start; i < nr_frags; i++) {
982 int j, newerr;
983 pending_ring_idx_t index;
984
985 pending_idx = (unsigned long)shinfo->frags[i].page;
986
987 /* Check error status: if okay then remember grant handle. */
988 newerr = (++gop)->status;
989 if (likely(!newerr)) {
990 /* Had a previous error? Invalidate this fragment. */
991 if (unlikely(err))
992 xen_netbk_idx_release(netbk, pending_idx);
993 continue;
994 }
995
996 /* Error on this fragment: respond to client with an error. */
997 txp = &netbk->pending_tx_info[pending_idx].req;
998 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
999 index = pending_index(netbk->pending_prod++);
1000 netbk->pending_ring[index] = pending_idx;
1001 xenvif_put(vif);
1002
1003 /* Not the first error? Preceding frags already invalidated. */
1004 if (err)
1005 continue;
1006
1007 /* First error: invalidate header and preceding fragments. */
1008 pending_idx = *((u16 *)skb->data);
1009 xen_netbk_idx_release(netbk, pending_idx);
1010 for (j = start; j < i; j++) {
1011 pending_idx = (unsigned long)shinfo->frags[i].page;
1012 xen_netbk_idx_release(netbk, pending_idx);
1013 }
1014
1015 /* Remember the error: invalidate all subsequent fragments. */
1016 err = newerr;
1017 }
1018
1019 *gopp = gop + 1;
1020 return err;
1021 }
1022
1023 static void xen_netbk_fill_frags(struct xen_netbk *netbk, struct sk_buff *skb)
1024 {
1025 struct skb_shared_info *shinfo = skb_shinfo(skb);
1026 int nr_frags = shinfo->nr_frags;
1027 int i;
1028
1029 for (i = 0; i < nr_frags; i++) {
1030 skb_frag_t *frag = shinfo->frags + i;
1031 struct xen_netif_tx_request *txp;
1032 unsigned long pending_idx;
1033
1034 pending_idx = (unsigned long)frag->page;
1035
1036 txp = &netbk->pending_tx_info[pending_idx].req;
1037 frag->page = virt_to_page(idx_to_kaddr(netbk, pending_idx));
1038 frag->size = txp->size;
1039 frag->page_offset = txp->offset;
1040
1041 skb->len += txp->size;
1042 skb->data_len += txp->size;
1043 skb->truesize += txp->size;
1044
1045 /* Take an extra reference to offset xen_netbk_idx_release */
1046 get_page(netbk->mmap_pages[pending_idx]);
1047 xen_netbk_idx_release(netbk, pending_idx);
1048 }
1049 }
1050
1051 static int xen_netbk_get_extras(struct xenvif *vif,
1052 struct xen_netif_extra_info *extras,
1053 int work_to_do)
1054 {
1055 struct xen_netif_extra_info extra;
1056 RING_IDX cons = vif->tx.req_cons;
1057
1058 do {
1059 if (unlikely(work_to_do-- <= 0)) {
1060 netdev_dbg(vif->dev, "Missing extra info\n");
1061 return -EBADR;
1062 }
1063
1064 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1065 sizeof(extra));
1066 if (unlikely(!extra.type ||
1067 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1068 vif->tx.req_cons = ++cons;
1069 netdev_dbg(vif->dev,
1070 "Invalid extra type: %d\n", extra.type);
1071 return -EINVAL;
1072 }
1073
1074 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1075 vif->tx.req_cons = ++cons;
1076 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1077
1078 return work_to_do;
1079 }
1080
1081 static int netbk_set_skb_gso(struct xenvif *vif,
1082 struct sk_buff *skb,
1083 struct xen_netif_extra_info *gso)
1084 {
1085 if (!gso->u.gso.size) {
1086 netdev_dbg(vif->dev, "GSO size must not be zero.\n");
1087 return -EINVAL;
1088 }
1089
1090 /* Currently only TCPv4 S.O. is supported. */
1091 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
1092 netdev_dbg(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1093 return -EINVAL;
1094 }
1095
1096 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1097 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1098
1099 /* Header must be checked, and gso_segs computed. */
1100 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1101 skb_shinfo(skb)->gso_segs = 0;
1102
1103 return 0;
1104 }
1105
1106 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1107 {
1108 struct iphdr *iph;
1109 unsigned char *th;
1110 int err = -EPROTO;
1111 int recalculate_partial_csum = 0;
1112
1113 /*
1114 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1115 * peers can fail to set NETRXF_csum_blank when sending a GSO
1116 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1117 * recalculate the partial checksum.
1118 */
1119 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1120 vif->rx_gso_checksum_fixup++;
1121 skb->ip_summed = CHECKSUM_PARTIAL;
1122 recalculate_partial_csum = 1;
1123 }
1124
1125 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1126 if (skb->ip_summed != CHECKSUM_PARTIAL)
1127 return 0;
1128
1129 if (skb->protocol != htons(ETH_P_IP))
1130 goto out;
1131
1132 iph = (void *)skb->data;
1133 th = skb->data + 4 * iph->ihl;
1134 if (th >= skb_tail_pointer(skb))
1135 goto out;
1136
1137 skb->csum_start = th - skb->head;
1138 switch (iph->protocol) {
1139 case IPPROTO_TCP:
1140 skb->csum_offset = offsetof(struct tcphdr, check);
1141
1142 if (recalculate_partial_csum) {
1143 struct tcphdr *tcph = (struct tcphdr *)th;
1144 tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1145 skb->len - iph->ihl*4,
1146 IPPROTO_TCP, 0);
1147 }
1148 break;
1149 case IPPROTO_UDP:
1150 skb->csum_offset = offsetof(struct udphdr, check);
1151
1152 if (recalculate_partial_csum) {
1153 struct udphdr *udph = (struct udphdr *)th;
1154 udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1155 skb->len - iph->ihl*4,
1156 IPPROTO_UDP, 0);
1157 }
1158 break;
1159 default:
1160 if (net_ratelimit())
1161 netdev_err(vif->dev,
1162 "Attempting to checksum a non-TCP/UDP packet, dropping a protocol %d packet\n",
1163 iph->protocol);
1164 goto out;
1165 }
1166
1167 if ((th + skb->csum_offset + 2) > skb_tail_pointer(skb))
1168 goto out;
1169
1170 err = 0;
1171
1172 out:
1173 return err;
1174 }
1175
1176 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1177 {
1178 unsigned long now = jiffies;
1179 unsigned long next_credit =
1180 vif->credit_timeout.expires +
1181 msecs_to_jiffies(vif->credit_usec / 1000);
1182
1183 /* Timer could already be pending in rare cases. */
1184 if (timer_pending(&vif->credit_timeout))
1185 return true;
1186
1187 /* Passed the point where we can replenish credit? */
1188 if (time_after_eq(now, next_credit)) {
1189 vif->credit_timeout.expires = now;
1190 tx_add_credit(vif);
1191 }
1192
1193 /* Still too big to send right now? Set a callback. */
1194 if (size > vif->remaining_credit) {
1195 vif->credit_timeout.data =
1196 (unsigned long)vif;
1197 vif->credit_timeout.function =
1198 tx_credit_callback;
1199 mod_timer(&vif->credit_timeout,
1200 next_credit);
1201
1202 return true;
1203 }
1204
1205 return false;
1206 }
1207
1208 static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
1209 {
1210 struct gnttab_copy *gop = netbk->tx_copy_ops, *request_gop;
1211 struct sk_buff *skb;
1212 int ret;
1213
1214 while (((nr_pending_reqs(netbk) + MAX_SKB_FRAGS) < MAX_PENDING_REQS) &&
1215 !list_empty(&netbk->net_schedule_list)) {
1216 struct xenvif *vif;
1217 struct xen_netif_tx_request txreq;
1218 struct xen_netif_tx_request txfrags[MAX_SKB_FRAGS];
1219 struct page *page;
1220 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1221 u16 pending_idx;
1222 RING_IDX idx;
1223 int work_to_do;
1224 unsigned int data_len;
1225 pending_ring_idx_t index;
1226
1227 /* Get a netif from the list with work to do. */
1228 vif = poll_net_schedule_list(netbk);
1229 if (!vif)
1230 continue;
1231
1232 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, work_to_do);
1233 if (!work_to_do) {
1234 xenvif_put(vif);
1235 continue;
1236 }
1237
1238 idx = vif->tx.req_cons;
1239 rmb(); /* Ensure that we see the request before we copy it. */
1240 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1241
1242 /* Credit-based scheduling. */
1243 if (txreq.size > vif->remaining_credit &&
1244 tx_credit_exceeded(vif, txreq.size)) {
1245 xenvif_put(vif);
1246 continue;
1247 }
1248
1249 vif->remaining_credit -= txreq.size;
1250
1251 work_to_do--;
1252 vif->tx.req_cons = ++idx;
1253
1254 memset(extras, 0, sizeof(extras));
1255 if (txreq.flags & XEN_NETTXF_extra_info) {
1256 work_to_do = xen_netbk_get_extras(vif, extras,
1257 work_to_do);
1258 idx = vif->tx.req_cons;
1259 if (unlikely(work_to_do < 0)) {
1260 netbk_tx_err(vif, &txreq, idx);
1261 continue;
1262 }
1263 }
1264
1265 ret = netbk_count_requests(vif, &txreq, txfrags, work_to_do);
1266 if (unlikely(ret < 0)) {
1267 netbk_tx_err(vif, &txreq, idx - ret);
1268 continue;
1269 }
1270 idx += ret;
1271
1272 if (unlikely(txreq.size < ETH_HLEN)) {
1273 netdev_dbg(vif->dev,
1274 "Bad packet size: %d\n", txreq.size);
1275 netbk_tx_err(vif, &txreq, idx);
1276 continue;
1277 }
1278
1279 /* No crossing a page as the payload mustn't fragment. */
1280 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1281 netdev_dbg(vif->dev,
1282 "txreq.offset: %x, size: %u, end: %lu\n",
1283 txreq.offset, txreq.size,
1284 (txreq.offset&~PAGE_MASK) + txreq.size);
1285 netbk_tx_err(vif, &txreq, idx);
1286 continue;
1287 }
1288
1289 index = pending_index(netbk->pending_cons);
1290 pending_idx = netbk->pending_ring[index];
1291
1292 data_len = (txreq.size > PKT_PROT_LEN &&
1293 ret < MAX_SKB_FRAGS) ?
1294 PKT_PROT_LEN : txreq.size;
1295
1296 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1297 GFP_ATOMIC | __GFP_NOWARN);
1298 if (unlikely(skb == NULL)) {
1299 netdev_dbg(vif->dev,
1300 "Can't allocate a skb in start_xmit.\n");
1301 netbk_tx_err(vif, &txreq, idx);
1302 break;
1303 }
1304
1305 /* Packets passed to netif_rx() must have some headroom. */
1306 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1307
1308 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1309 struct xen_netif_extra_info *gso;
1310 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1311
1312 if (netbk_set_skb_gso(vif, skb, gso)) {
1313 kfree_skb(skb);
1314 netbk_tx_err(vif, &txreq, idx);
1315 continue;
1316 }
1317 }
1318
1319 /* XXX could copy straight to head */
1320 page = xen_netbk_alloc_page(netbk, skb, pending_idx);
1321 if (!page) {
1322 kfree_skb(skb);
1323 netbk_tx_err(vif, &txreq, idx);
1324 continue;
1325 }
1326
1327 netbk->mmap_pages[pending_idx] = page;
1328
1329 gop->source.u.ref = txreq.gref;
1330 gop->source.domid = vif->domid;
1331 gop->source.offset = txreq.offset;
1332
1333 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1334 gop->dest.domid = DOMID_SELF;
1335 gop->dest.offset = txreq.offset;
1336
1337 gop->len = txreq.size;
1338 gop->flags = GNTCOPY_source_gref;
1339
1340 gop++;
1341
1342 memcpy(&netbk->pending_tx_info[pending_idx].req,
1343 &txreq, sizeof(txreq));
1344 netbk->pending_tx_info[pending_idx].vif = vif;
1345 *((u16 *)skb->data) = pending_idx;
1346
1347 __skb_put(skb, data_len);
1348
1349 skb_shinfo(skb)->nr_frags = ret;
1350 if (data_len < txreq.size) {
1351 skb_shinfo(skb)->nr_frags++;
1352 skb_shinfo(skb)->frags[0].page =
1353 (void *)(unsigned long)pending_idx;
1354 } else {
1355 /* Discriminate from any valid pending_idx value. */
1356 skb_shinfo(skb)->frags[0].page = (void *)~0UL;
1357 }
1358
1359 __skb_queue_tail(&netbk->tx_queue, skb);
1360
1361 netbk->pending_cons++;
1362
1363 request_gop = xen_netbk_get_requests(netbk, vif,
1364 skb, txfrags, gop);
1365 if (request_gop == NULL) {
1366 kfree_skb(skb);
1367 netbk_tx_err(vif, &txreq, idx);
1368 continue;
1369 }
1370 gop = request_gop;
1371
1372 vif->tx.req_cons = idx;
1373 xen_netbk_check_rx_xenvif(vif);
1374
1375 if ((gop-netbk->tx_copy_ops) >= ARRAY_SIZE(netbk->tx_copy_ops))
1376 break;
1377 }
1378
1379 return gop - netbk->tx_copy_ops;
1380 }
1381
1382 static void xen_netbk_tx_submit(struct xen_netbk *netbk)
1383 {
1384 struct gnttab_copy *gop = netbk->tx_copy_ops;
1385 struct sk_buff *skb;
1386
1387 while ((skb = __skb_dequeue(&netbk->tx_queue)) != NULL) {
1388 struct xen_netif_tx_request *txp;
1389 struct xenvif *vif;
1390 u16 pending_idx;
1391 unsigned data_len;
1392
1393 pending_idx = *((u16 *)skb->data);
1394 vif = netbk->pending_tx_info[pending_idx].vif;
1395 txp = &netbk->pending_tx_info[pending_idx].req;
1396
1397 /* Check the remap error code. */
1398 if (unlikely(xen_netbk_tx_check_gop(netbk, skb, &gop))) {
1399 netdev_dbg(vif->dev, "netback grant failed.\n");
1400 skb_shinfo(skb)->nr_frags = 0;
1401 kfree_skb(skb);
1402 continue;
1403 }
1404
1405 data_len = skb->len;
1406 memcpy(skb->data,
1407 (void *)(idx_to_kaddr(netbk, pending_idx)|txp->offset),
1408 data_len);
1409 if (data_len < txp->size) {
1410 /* Append the packet payload as a fragment. */
1411 txp->offset += data_len;
1412 txp->size -= data_len;
1413 } else {
1414 /* Schedule a response immediately. */
1415 xen_netbk_idx_release(netbk, pending_idx);
1416 }
1417
1418 if (txp->flags & XEN_NETTXF_csum_blank)
1419 skb->ip_summed = CHECKSUM_PARTIAL;
1420 else if (txp->flags & XEN_NETTXF_data_validated)
1421 skb->ip_summed = CHECKSUM_UNNECESSARY;
1422
1423 xen_netbk_fill_frags(netbk, skb);
1424
1425 /*
1426 * If the initial fragment was < PKT_PROT_LEN then
1427 * pull through some bytes from the other fragments to
1428 * increase the linear region to PKT_PROT_LEN bytes.
1429 */
1430 if (skb_headlen(skb) < PKT_PROT_LEN && skb_is_nonlinear(skb)) {
1431 int target = min_t(int, skb->len, PKT_PROT_LEN);
1432 __pskb_pull_tail(skb, target - skb_headlen(skb));
1433 }
1434
1435 skb->dev = vif->dev;
1436 skb->protocol = eth_type_trans(skb, skb->dev);
1437
1438 if (checksum_setup(vif, skb)) {
1439 netdev_dbg(vif->dev,
1440 "Can't setup checksum in net_tx_action\n");
1441 kfree_skb(skb);
1442 continue;
1443 }
1444
1445 vif->dev->stats.rx_bytes += skb->len;
1446 vif->dev->stats.rx_packets++;
1447
1448 xenvif_receive_skb(vif, skb);
1449 }
1450 }
1451
1452 /* Called after netfront has transmitted */
1453 static void xen_netbk_tx_action(struct xen_netbk *netbk)
1454 {
1455 unsigned nr_gops;
1456 int ret;
1457
1458 nr_gops = xen_netbk_tx_build_gops(netbk);
1459
1460 if (nr_gops == 0)
1461 return;
1462 ret = HYPERVISOR_grant_table_op(GNTTABOP_copy,
1463 netbk->tx_copy_ops, nr_gops);
1464 BUG_ON(ret);
1465
1466 xen_netbk_tx_submit(netbk);
1467
1468 }
1469
1470 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx)
1471 {
1472 struct xenvif *vif;
1473 struct pending_tx_info *pending_tx_info;
1474 pending_ring_idx_t index;
1475
1476 /* Already complete? */
1477 if (netbk->mmap_pages[pending_idx] == NULL)
1478 return;
1479
1480 pending_tx_info = &netbk->pending_tx_info[pending_idx];
1481
1482 vif = pending_tx_info->vif;
1483
1484 make_tx_response(vif, &pending_tx_info->req, XEN_NETIF_RSP_OKAY);
1485
1486 index = pending_index(netbk->pending_prod++);
1487 netbk->pending_ring[index] = pending_idx;
1488
1489 xenvif_put(vif);
1490
1491 netbk->mmap_pages[pending_idx]->mapping = 0;
1492 put_page(netbk->mmap_pages[pending_idx]);
1493 netbk->mmap_pages[pending_idx] = NULL;
1494 }
1495
1496 static void make_tx_response(struct xenvif *vif,
1497 struct xen_netif_tx_request *txp,
1498 s8 st)
1499 {
1500 RING_IDX i = vif->tx.rsp_prod_pvt;
1501 struct xen_netif_tx_response *resp;
1502 int notify;
1503
1504 resp = RING_GET_RESPONSE(&vif->tx, i);
1505 resp->id = txp->id;
1506 resp->status = st;
1507
1508 if (txp->flags & XEN_NETTXF_extra_info)
1509 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1510
1511 vif->tx.rsp_prod_pvt = ++i;
1512 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1513 if (notify)
1514 notify_remote_via_irq(vif->irq);
1515 }
1516
1517 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1518 u16 id,
1519 s8 st,
1520 u16 offset,
1521 u16 size,
1522 u16 flags)
1523 {
1524 RING_IDX i = vif->rx.rsp_prod_pvt;
1525 struct xen_netif_rx_response *resp;
1526
1527 resp = RING_GET_RESPONSE(&vif->rx, i);
1528 resp->offset = offset;
1529 resp->flags = flags;
1530 resp->id = id;
1531 resp->status = (s16)size;
1532 if (st < 0)
1533 resp->status = (s16)st;
1534
1535 vif->rx.rsp_prod_pvt = ++i;
1536
1537 return resp;
1538 }
1539
1540 static inline int rx_work_todo(struct xen_netbk *netbk)
1541 {
1542 return !skb_queue_empty(&netbk->rx_queue);
1543 }
1544
1545 static inline int tx_work_todo(struct xen_netbk *netbk)
1546 {
1547
1548 if (((nr_pending_reqs(netbk) + MAX_SKB_FRAGS) < MAX_PENDING_REQS) &&
1549 !list_empty(&netbk->net_schedule_list))
1550 return 1;
1551
1552 return 0;
1553 }
1554
1555 static int xen_netbk_kthread(void *data)
1556 {
1557 struct xen_netbk *netbk = data;
1558 while (!kthread_should_stop()) {
1559 wait_event_interruptible(netbk->wq,
1560 rx_work_todo(netbk) ||
1561 tx_work_todo(netbk) ||
1562 kthread_should_stop());
1563 cond_resched();
1564
1565 if (kthread_should_stop())
1566 break;
1567
1568 if (rx_work_todo(netbk))
1569 xen_netbk_rx_action(netbk);
1570
1571 if (tx_work_todo(netbk))
1572 xen_netbk_tx_action(netbk);
1573 }
1574
1575 return 0;
1576 }
1577
1578 void xen_netbk_unmap_frontend_rings(struct xenvif *vif)
1579 {
1580 struct gnttab_unmap_grant_ref op;
1581
1582 if (vif->tx.sring) {
1583 gnttab_set_unmap_op(&op, (unsigned long)vif->tx_comms_area->addr,
1584 GNTMAP_host_map, vif->tx_shmem_handle);
1585
1586 if (HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref, &op, 1))
1587 BUG();
1588 }
1589
1590 if (vif->rx.sring) {
1591 gnttab_set_unmap_op(&op, (unsigned long)vif->rx_comms_area->addr,
1592 GNTMAP_host_map, vif->rx_shmem_handle);
1593
1594 if (HYPERVISOR_grant_table_op(GNTTABOP_unmap_grant_ref, &op, 1))
1595 BUG();
1596 }
1597 if (vif->rx_comms_area)
1598 free_vm_area(vif->rx_comms_area);
1599 if (vif->tx_comms_area)
1600 free_vm_area(vif->tx_comms_area);
1601 }
1602
1603 int xen_netbk_map_frontend_rings(struct xenvif *vif,
1604 grant_ref_t tx_ring_ref,
1605 grant_ref_t rx_ring_ref)
1606 {
1607 struct gnttab_map_grant_ref op;
1608 struct xen_netif_tx_sring *txs;
1609 struct xen_netif_rx_sring *rxs;
1610
1611 int err = -ENOMEM;
1612
1613 vif->tx_comms_area = alloc_vm_area(PAGE_SIZE);
1614 if (vif->tx_comms_area == NULL)
1615 goto err;
1616
1617 vif->rx_comms_area = alloc_vm_area(PAGE_SIZE);
1618 if (vif->rx_comms_area == NULL)
1619 goto err;
1620
1621 gnttab_set_map_op(&op, (unsigned long)vif->tx_comms_area->addr,
1622 GNTMAP_host_map, tx_ring_ref, vif->domid);
1623
1624 if (HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, &op, 1))
1625 BUG();
1626
1627 if (op.status) {
1628 netdev_warn(vif->dev,
1629 "failed to map tx ring. err=%d status=%d\n",
1630 err, op.status);
1631 err = op.status;
1632 goto err;
1633 }
1634
1635 vif->tx_shmem_ref = tx_ring_ref;
1636 vif->tx_shmem_handle = op.handle;
1637
1638 txs = (struct xen_netif_tx_sring *)vif->tx_comms_area->addr;
1639 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1640
1641 gnttab_set_map_op(&op, (unsigned long)vif->rx_comms_area->addr,
1642 GNTMAP_host_map, rx_ring_ref, vif->domid);
1643
1644 if (HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, &op, 1))
1645 BUG();
1646
1647 if (op.status) {
1648 netdev_warn(vif->dev,
1649 "failed to map rx ring. err=%d status=%d\n",
1650 err, op.status);
1651 err = op.status;
1652 goto err;
1653 }
1654
1655 vif->rx_shmem_ref = rx_ring_ref;
1656 vif->rx_shmem_handle = op.handle;
1657 vif->rx_req_cons_peek = 0;
1658
1659 rxs = (struct xen_netif_rx_sring *)vif->rx_comms_area->addr;
1660 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1661
1662 return 0;
1663
1664 err:
1665 xen_netbk_unmap_frontend_rings(vif);
1666 return err;
1667 }
1668
1669 static int __init netback_init(void)
1670 {
1671 int i;
1672 int rc = 0;
1673 int group;
1674
1675 if (!xen_pv_domain())
1676 return -ENODEV;
1677
1678 xen_netbk_group_nr = num_online_cpus();
1679 xen_netbk = vzalloc(sizeof(struct xen_netbk) * xen_netbk_group_nr);
1680 if (!xen_netbk) {
1681 printk(KERN_ALERT "%s: out of memory\n", __func__);
1682 return -ENOMEM;
1683 }
1684
1685 for (group = 0; group < xen_netbk_group_nr; group++) {
1686 struct xen_netbk *netbk = &xen_netbk[group];
1687 skb_queue_head_init(&netbk->rx_queue);
1688 skb_queue_head_init(&netbk->tx_queue);
1689
1690 init_timer(&netbk->net_timer);
1691 netbk->net_timer.data = (unsigned long)netbk;
1692 netbk->net_timer.function = xen_netbk_alarm;
1693
1694 netbk->pending_cons = 0;
1695 netbk->pending_prod = MAX_PENDING_REQS;
1696 for (i = 0; i < MAX_PENDING_REQS; i++)
1697 netbk->pending_ring[i] = i;
1698
1699 init_waitqueue_head(&netbk->wq);
1700 netbk->task = kthread_create(xen_netbk_kthread,
1701 (void *)netbk,
1702 "netback/%u", group);
1703
1704 if (IS_ERR(netbk->task)) {
1705 printk(KERN_ALERT "kthread_run() fails at netback\n");
1706 del_timer(&netbk->net_timer);
1707 rc = PTR_ERR(netbk->task);
1708 goto failed_init;
1709 }
1710
1711 kthread_bind(netbk->task, group);
1712
1713 INIT_LIST_HEAD(&netbk->net_schedule_list);
1714
1715 spin_lock_init(&netbk->net_schedule_list_lock);
1716
1717 atomic_set(&netbk->netfront_count, 0);
1718
1719 wake_up_process(netbk->task);
1720 }
1721
1722 rc = xenvif_xenbus_init();
1723 if (rc)
1724 goto failed_init;
1725
1726 return 0;
1727
1728 failed_init:
1729 while (--group >= 0) {
1730 struct xen_netbk *netbk = &xen_netbk[group];
1731 for (i = 0; i < MAX_PENDING_REQS; i++) {
1732 if (netbk->mmap_pages[i])
1733 __free_page(netbk->mmap_pages[i]);
1734 }
1735 del_timer(&netbk->net_timer);
1736 kthread_stop(netbk->task);
1737 }
1738 vfree(xen_netbk);
1739 return rc;
1740
1741 }
1742
1743 module_init(netback_init);
1744
1745 MODULE_LICENSE("Dual BSD/GPL");
linux-next