?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 3.4.102
[linux/fpc-iii.git] / drivers / net / xen-netback / netback.c
blobcd4966bcb2d264f1d5e838846cae0001187c2b1f
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 /*
50 * This is the maximum slots a skb can have. If a guest sends a skb
51 * which exceeds this limit it is considered malicious.
52 */
53 #define MAX_SKB_SLOTS_DEFAULT 20
54 static unsigned int max_skb_slots = MAX_SKB_SLOTS_DEFAULT;
55 module_param(max_skb_slots, uint, 0444);
56
57 typedef unsigned int pending_ring_idx_t;
58 #define INVALID_PENDING_RING_IDX (~0U)
59
60 struct pending_tx_info {
61 struct xen_netif_tx_request req; /* coalesced tx request */
62 struct xenvif *vif;
63 pending_ring_idx_t head; /* head != INVALID_PENDING_RING_IDX
64 * if it is head of one or more tx
65 * reqs
66 */
67 };
68
69 struct netbk_rx_meta {
70 int id;
71 int size;
72 int gso_size;
73 };
74
75 #define MAX_PENDING_REQS 256
76
77 /* Discriminate from any valid pending_idx value. */
78 #define INVALID_PENDING_IDX 0xFFFF
79
80 #define MAX_BUFFER_OFFSET PAGE_SIZE
81
82 /* extra field used in struct page */
83 union page_ext {
84 struct {
85 #if BITS_PER_LONG < 64
86 #define IDX_WIDTH 8
87 #define GROUP_WIDTH (BITS_PER_LONG - IDX_WIDTH)
88 unsigned int group:GROUP_WIDTH;
89 unsigned int idx:IDX_WIDTH;
90 #else
91 unsigned int group, idx;
92 #endif
93 } e;
94 void *mapping;
95 };
96
97 struct xen_netbk {
98 wait_queue_head_t wq;
99 struct task_struct *task;
100
101 struct sk_buff_head rx_queue;
102 struct sk_buff_head tx_queue;
103
104 struct timer_list net_timer;
105
106 struct page *mmap_pages[MAX_PENDING_REQS];
107
108 pending_ring_idx_t pending_prod;
109 pending_ring_idx_t pending_cons;
110 struct list_head net_schedule_list;
111
112 /* Protect the net_schedule_list in netif. */
113 spinlock_t net_schedule_list_lock;
114
115 atomic_t netfront_count;
116
117 struct pending_tx_info pending_tx_info[MAX_PENDING_REQS];
118 /* Coalescing tx requests before copying makes number of grant
119 * copy ops greater or equal to number of slots required. In
120 * worst case a tx request consumes 2 gnttab_copy.
121 */
122 struct gnttab_copy tx_copy_ops[2*MAX_PENDING_REQS];
123
124 u16 pending_ring[MAX_PENDING_REQS];
125
126 /*
127 * Given MAX_BUFFER_OFFSET of 4096 the worst case is that each
128 * head/fragment page uses 2 copy operations because it
129 * straddles two buffers in the frontend.
130 */
131 struct gnttab_copy grant_copy_op[2*XEN_NETIF_RX_RING_SIZE];
132 struct netbk_rx_meta meta[2*XEN_NETIF_RX_RING_SIZE];
133 };
134
135 static struct xen_netbk *xen_netbk;
136 static int xen_netbk_group_nr;
137
138 /*
139 * If head != INVALID_PENDING_RING_IDX, it means this tx request is head of
140 * one or more merged tx requests, otherwise it is the continuation of
141 * previous tx request.
142 */
143 static inline int pending_tx_is_head(struct xen_netbk *netbk, RING_IDX idx)
144 {
145 return netbk->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
146 }
147
148 void xen_netbk_add_xenvif(struct xenvif *vif)
149 {
150 int i;
151 int min_netfront_count;
152 int min_group = 0;
153 struct xen_netbk *netbk;
154
155 min_netfront_count = atomic_read(&xen_netbk[0].netfront_count);
156 for (i = 0; i < xen_netbk_group_nr; i++) {
157 int netfront_count = atomic_read(&xen_netbk[i].netfront_count);
158 if (netfront_count < min_netfront_count) {
159 min_group = i;
160 min_netfront_count = netfront_count;
161 }
162 }
163
164 netbk = &xen_netbk[min_group];
165
166 vif->netbk = netbk;
167 atomic_inc(&netbk->netfront_count);
168 }
169
170 void xen_netbk_remove_xenvif(struct xenvif *vif)
171 {
172 struct xen_netbk *netbk = vif->netbk;
173 vif->netbk = NULL;
174 atomic_dec(&netbk->netfront_count);
175 }
176
177 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
178 u8 status);
179 static void make_tx_response(struct xenvif *vif,
180 struct xen_netif_tx_request *txp,
181 s8 st);
182 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
183 u16 id,
184 s8 st,
185 u16 offset,
186 u16 size,
187 u16 flags);
188
189 static inline unsigned long idx_to_pfn(struct xen_netbk *netbk,
190 u16 idx)
191 {
192 return page_to_pfn(netbk->mmap_pages[idx]);
193 }
194
195 static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
196 u16 idx)
197 {
198 return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
199 }
200
201 /* extra field used in struct page */
202 static inline void set_page_ext(struct page *pg, struct xen_netbk *netbk,
203 unsigned int idx)
204 {
205 unsigned int group = netbk - xen_netbk;
206 union page_ext ext = { .e = { .group = group + 1, .idx = idx } };
207
208 BUILD_BUG_ON(sizeof(ext) > sizeof(ext.mapping));
209 pg->mapping = ext.mapping;
210 }
211
212 static int get_page_ext(struct page *pg,
213 unsigned int *pgroup, unsigned int *pidx)
214 {
215 union page_ext ext = { .mapping = pg->mapping };
216 struct xen_netbk *netbk;
217 unsigned int group, idx;
218
219 group = ext.e.group - 1;
220
221 if (group < 0 || group >= xen_netbk_group_nr)
222 return 0;
223
224 netbk = &xen_netbk[group];
225
226 idx = ext.e.idx;
227
228 if ((idx < 0) || (idx >= MAX_PENDING_REQS))
229 return 0;
230
231 if (netbk->mmap_pages[idx] != pg)
232 return 0;
233
234 *pgroup = group;
235 *pidx = idx;
236
237 return 1;
238 }
239
240 /*
241 * This is the amount of packet we copy rather than map, so that the
242 * guest can't fiddle with the contents of the headers while we do
243 * packet processing on them (netfilter, routing, etc).
244 */
245 #define PKT_PROT_LEN (ETH_HLEN + \
246 VLAN_HLEN + \
247 sizeof(struct iphdr) + MAX_IPOPTLEN + \
248 sizeof(struct tcphdr) + MAX_TCP_OPTION_SPACE)
249
250 static u16 frag_get_pending_idx(skb_frag_t *frag)
251 {
252 return (u16)frag->page_offset;
253 }
254
255 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
256 {
257 frag->page_offset = pending_idx;
258 }
259
260 static inline pending_ring_idx_t pending_index(unsigned i)
261 {
262 return i & (MAX_PENDING_REQS-1);
263 }
264
265 static inline pending_ring_idx_t nr_pending_reqs(struct xen_netbk *netbk)
266 {
267 return MAX_PENDING_REQS -
268 netbk->pending_prod + netbk->pending_cons;
269 }
270
271 static void xen_netbk_kick_thread(struct xen_netbk *netbk)
272 {
273 wake_up(&netbk->wq);
274 }
275
276 static int max_required_rx_slots(struct xenvif *vif)
277 {
278 int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
279
280 /* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
281 if (vif->can_sg || vif->gso || vif->gso_prefix)
282 max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
283
284 return max;
285 }
286
287 int xen_netbk_rx_ring_full(struct xenvif *vif)
288 {
289 RING_IDX peek = vif->rx_req_cons_peek;
290 RING_IDX needed = max_required_rx_slots(vif);
291
292 return ((vif->rx.sring->req_prod - peek) < needed) ||
293 ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
294 }
295
296 int xen_netbk_must_stop_queue(struct xenvif *vif)
297 {
298 if (!xen_netbk_rx_ring_full(vif))
299 return 0;
300
301 vif->rx.sring->req_event = vif->rx_req_cons_peek +
302 max_required_rx_slots(vif);
303 mb(); /* request notification /then/ check the queue */
304
305 return xen_netbk_rx_ring_full(vif);
306 }
307
308 /*
309 * Returns true if we should start a new receive buffer instead of
310 * adding 'size' bytes to a buffer which currently contains 'offset'
311 * bytes.
312 */
313 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
314 {
315 /* simple case: we have completely filled the current buffer. */
316 if (offset == MAX_BUFFER_OFFSET)
317 return true;
318
319 /*
320 * complex case: start a fresh buffer if the current frag
321 * would overflow the current buffer but only if:
322 * (i) this frag would fit completely in the next buffer
323 * and (ii) there is already some data in the current buffer
324 * and (iii) this is not the head buffer.
325 *
326 * Where:
327 * - (i) stops us splitting a frag into two copies
328 * unless the frag is too large for a single buffer.
329 * - (ii) stops us from leaving a buffer pointlessly empty.
330 * - (iii) stops us leaving the first buffer
331 * empty. Strictly speaking this is already covered
332 * by (ii) but is explicitly checked because
333 * netfront relies on the first buffer being
334 * non-empty and can crash otherwise.
335 *
336 * This means we will effectively linearise small
337 * frags but do not needlessly split large buffers
338 * into multiple copies tend to give large frags their
339 * own buffers as before.
340 */
341 BUG_ON(size > MAX_BUFFER_OFFSET);
342 if ((offset + size > MAX_BUFFER_OFFSET) && offset && !head)
343 return true;
344
345 return false;
346 }
347
348 /*
349 * Figure out how many ring slots we're going to need to send @skb to
350 * the guest. This function is essentially a dry run of
351 * netbk_gop_frag_copy.
352 */
353 unsigned int xen_netbk_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
354 {
355 unsigned int count;
356 int i, copy_off;
357
358 count = DIV_ROUND_UP(
359 offset_in_page(skb->data)+skb_headlen(skb), PAGE_SIZE);
360
361 copy_off = skb_headlen(skb) % PAGE_SIZE;
362
363 if (skb_shinfo(skb)->gso_size)
364 count++;
365
366 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
367 unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
368 unsigned long bytes;
369 while (size > 0) {
370 BUG_ON(copy_off > MAX_BUFFER_OFFSET);
371
372 if (start_new_rx_buffer(copy_off, size, 0)) {
373 count++;
374 copy_off = 0;
375 }
376
377 bytes = size;
378 if (copy_off + bytes > MAX_BUFFER_OFFSET)
379 bytes = MAX_BUFFER_OFFSET - copy_off;
380
381 copy_off += bytes;
382 size -= bytes;
383 }
384 }
385 return count;
386 }
387
388 struct netrx_pending_operations {
389 unsigned copy_prod, copy_cons;
390 unsigned meta_prod, meta_cons;
391 struct gnttab_copy *copy;
392 struct netbk_rx_meta *meta;
393 int copy_off;
394 grant_ref_t copy_gref;
395 };
396
397 static struct netbk_rx_meta *get_next_rx_buffer(struct xenvif *vif,
398 struct netrx_pending_operations *npo)
399 {
400 struct netbk_rx_meta *meta;
401 struct xen_netif_rx_request *req;
402
403 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
404
405 meta = npo->meta + npo->meta_prod++;
406 meta->gso_size = 0;
407 meta->size = 0;
408 meta->id = req->id;
409
410 npo->copy_off = 0;
411 npo->copy_gref = req->gref;
412
413 return meta;
414 }
415
416 /*
417 * Set up the grant operations for this fragment. If it's a flipping
418 * interface, we also set up the unmap request from here.
419 */
420 static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
421 struct netrx_pending_operations *npo,
422 struct page *page, unsigned long size,
423 unsigned long offset, int *head)
424 {
425 struct gnttab_copy *copy_gop;
426 struct netbk_rx_meta *meta;
427 /*
428 * These variables are used iff get_page_ext returns true,
429 * in which case they are guaranteed to be initialized.
430 */
431 unsigned int uninitialized_var(group), uninitialized_var(idx);
432 int foreign = get_page_ext(page, &group, &idx);
433 unsigned long bytes;
434
435 /* Data must not cross a page boundary. */
436 BUG_ON(size + offset > PAGE_SIZE);
437
438 meta = npo->meta + npo->meta_prod - 1;
439
440 while (size > 0) {
441 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
442
443 if (start_new_rx_buffer(npo->copy_off, size, *head)) {
444 /*
445 * Netfront requires there to be some data in the head
446 * buffer.
447 */
448 BUG_ON(*head);
449
450 meta = get_next_rx_buffer(vif, npo);
451 }
452
453 bytes = size;
454 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
455 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
456
457 copy_gop = npo->copy + npo->copy_prod++;
458 copy_gop->flags = GNTCOPY_dest_gref;
459 if (foreign) {
460 struct xen_netbk *netbk = &xen_netbk[group];
461 struct pending_tx_info *src_pend;
462
463 src_pend = &netbk->pending_tx_info[idx];
464
465 copy_gop->source.domid = src_pend->vif->domid;
466 copy_gop->source.u.ref = src_pend->req.gref;
467 copy_gop->flags |= GNTCOPY_source_gref;
468 } else {
469 void *vaddr = page_address(page);
470 copy_gop->source.domid = DOMID_SELF;
471 copy_gop->source.u.gmfn = virt_to_mfn(vaddr);
472 }
473 copy_gop->source.offset = offset;
474 copy_gop->dest.domid = vif->domid;
475
476 copy_gop->dest.offset = npo->copy_off;
477 copy_gop->dest.u.ref = npo->copy_gref;
478 copy_gop->len = bytes;
479
480 npo->copy_off += bytes;
481 meta->size += bytes;
482
483 offset += bytes;
484 size -= bytes;
485
486 /* Leave a gap for the GSO descriptor. */
487 if (*head && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
488 vif->rx.req_cons++;
489
490 *head = 0; /* There must be something in this buffer now. */
491
492 }
493 }
494
495 /*
496 * Prepare an SKB to be transmitted to the frontend.
497 *
498 * This function is responsible for allocating grant operations, meta
499 * structures, etc.
500 *
501 * It returns the number of meta structures consumed. The number of
502 * ring slots used is always equal to the number of meta slots used
503 * plus the number of GSO descriptors used. Currently, we use either
504 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
505 * frontend-side LRO).
506 */
507 static int netbk_gop_skb(struct sk_buff *skb,
508 struct netrx_pending_operations *npo)
509 {
510 struct xenvif *vif = netdev_priv(skb->dev);
511 int nr_frags = skb_shinfo(skb)->nr_frags;
512 int i;
513 struct xen_netif_rx_request *req;
514 struct netbk_rx_meta *meta;
515 unsigned char *data;
516 int head = 1;
517 int old_meta_prod;
518
519 old_meta_prod = npo->meta_prod;
520
521 /* Set up a GSO prefix descriptor, if necessary */
522 if (skb_shinfo(skb)->gso_size && vif->gso_prefix) {
523 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
524 meta = npo->meta + npo->meta_prod++;
525 meta->gso_size = skb_shinfo(skb)->gso_size;
526 meta->size = 0;
527 meta->id = req->id;
528 }
529
530 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
531 meta = npo->meta + npo->meta_prod++;
532
533 if (!vif->gso_prefix)
534 meta->gso_size = skb_shinfo(skb)->gso_size;
535 else
536 meta->gso_size = 0;
537
538 meta->size = 0;
539 meta->id = req->id;
540 npo->copy_off = 0;
541 npo->copy_gref = req->gref;
542
543 data = skb->data;
544 while (data < skb_tail_pointer(skb)) {
545 unsigned int offset = offset_in_page(data);
546 unsigned int len = PAGE_SIZE - offset;
547
548 if (data + len > skb_tail_pointer(skb))
549 len = skb_tail_pointer(skb) - data;
550
551 netbk_gop_frag_copy(vif, skb, npo,
552 virt_to_page(data), len, offset, &head);
553 data += len;
554 }
555
556 for (i = 0; i < nr_frags; i++) {
557 netbk_gop_frag_copy(vif, skb, npo,
558 skb_frag_page(&skb_shinfo(skb)->frags[i]),
559 skb_frag_size(&skb_shinfo(skb)->frags[i]),
560 skb_shinfo(skb)->frags[i].page_offset,
561 &head);
562 }
563
564 return npo->meta_prod - old_meta_prod;
565 }
566
567 /*
568 * This is a twin to netbk_gop_skb. Assume that netbk_gop_skb was
569 * used to set up the operations on the top of
570 * netrx_pending_operations, which have since been done. Check that
571 * they didn't give any errors and advance over them.
572 */
573 static int netbk_check_gop(struct xenvif *vif, int nr_meta_slots,
574 struct netrx_pending_operations *npo)
575 {
576 struct gnttab_copy *copy_op;
577 int status = XEN_NETIF_RSP_OKAY;
578 int i;
579
580 for (i = 0; i < nr_meta_slots; i++) {
581 copy_op = npo->copy + npo->copy_cons++;
582 if (copy_op->status != GNTST_okay) {
583 netdev_dbg(vif->dev,
584 "Bad status %d from copy to DOM%d.\n",
585 copy_op->status, vif->domid);
586 status = XEN_NETIF_RSP_ERROR;
587 }
588 }
589
590 return status;
591 }
592
593 static void netbk_add_frag_responses(struct xenvif *vif, int status,
594 struct netbk_rx_meta *meta,
595 int nr_meta_slots)
596 {
597 int i;
598 unsigned long offset;
599
600 /* No fragments used */
601 if (nr_meta_slots <= 1)
602 return;
603
604 nr_meta_slots--;
605
606 for (i = 0; i < nr_meta_slots; i++) {
607 int flags;
608 if (i == nr_meta_slots - 1)
609 flags = 0;
610 else
611 flags = XEN_NETRXF_more_data;
612
613 offset = 0;
614 make_rx_response(vif, meta[i].id, status, offset,
615 meta[i].size, flags);
616 }
617 }
618
619 struct skb_cb_overlay {
620 int meta_slots_used;
621 };
622
623 static void xen_netbk_rx_action(struct xen_netbk *netbk)
624 {
625 struct xenvif *vif = NULL, *tmp;
626 s8 status;
627 u16 irq, flags;
628 struct xen_netif_rx_response *resp;
629 struct sk_buff_head rxq;
630 struct sk_buff *skb;
631 LIST_HEAD(notify);
632 int ret;
633 int nr_frags;
634 int count;
635 unsigned long offset;
636 struct skb_cb_overlay *sco;
637
638 struct netrx_pending_operations npo = {
639 .copy = netbk->grant_copy_op,
640 .meta = netbk->meta,
641 };
642
643 skb_queue_head_init(&rxq);
644
645 count = 0;
646
647 while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
648 vif = netdev_priv(skb->dev);
649 nr_frags = skb_shinfo(skb)->nr_frags;
650
651 sco = (struct skb_cb_overlay *)skb->cb;
652 sco->meta_slots_used = netbk_gop_skb(skb, &npo);
653
654 count += nr_frags + 1;
655
656 __skb_queue_tail(&rxq, skb);
657
658 /* Filled the batch queue? */
659 /* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
660 if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
661 break;
662 }
663
664 BUG_ON(npo.meta_prod > ARRAY_SIZE(netbk->meta));
665
666 if (!npo.copy_prod)
667 return;
668
669 BUG_ON(npo.copy_prod > ARRAY_SIZE(netbk->grant_copy_op));
670 ret = HYPERVISOR_grant_table_op(GNTTABOP_copy, &netbk->grant_copy_op,
671 npo.copy_prod);
672 BUG_ON(ret != 0);
673
674 while ((skb = __skb_dequeue(&rxq)) != NULL) {
675 sco = (struct skb_cb_overlay *)skb->cb;
676
677 vif = netdev_priv(skb->dev);
678
679 if (netbk->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
680 resp = RING_GET_RESPONSE(&vif->rx,
681 vif->rx.rsp_prod_pvt++);
682
683 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
684
685 resp->offset = netbk->meta[npo.meta_cons].gso_size;
686 resp->id = netbk->meta[npo.meta_cons].id;
687 resp->status = sco->meta_slots_used;
688
689 npo.meta_cons++;
690 sco->meta_slots_used--;
691 }
692
693
694 vif->dev->stats.tx_bytes += skb->len;
695 vif->dev->stats.tx_packets++;
696
697 status = netbk_check_gop(vif, sco->meta_slots_used, &npo);
698
699 if (sco->meta_slots_used == 1)
700 flags = 0;
701 else
702 flags = XEN_NETRXF_more_data;
703
704 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
705 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
706 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
707 /* remote but checksummed. */
708 flags |= XEN_NETRXF_data_validated;
709
710 offset = 0;
711 resp = make_rx_response(vif, netbk->meta[npo.meta_cons].id,
712 status, offset,
713 netbk->meta[npo.meta_cons].size,
714 flags);
715
716 if (netbk->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
717 struct xen_netif_extra_info *gso =
718 (struct xen_netif_extra_info *)
719 RING_GET_RESPONSE(&vif->rx,
720 vif->rx.rsp_prod_pvt++);
721
722 resp->flags |= XEN_NETRXF_extra_info;
723
724 gso->u.gso.size = netbk->meta[npo.meta_cons].gso_size;
725 gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
726 gso->u.gso.pad = 0;
727 gso->u.gso.features = 0;
728
729 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
730 gso->flags = 0;
731 }
732
733 netbk_add_frag_responses(vif, status,
734 netbk->meta + npo.meta_cons + 1,
735 sco->meta_slots_used);
736
737 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
738 irq = vif->irq;
739 if (ret && list_empty(&vif->notify_list))
740 list_add_tail(&vif->notify_list, &notify);
741
742 xenvif_notify_tx_completion(vif);
743
744 xenvif_put(vif);
745 npo.meta_cons += sco->meta_slots_used;
746 dev_kfree_skb(skb);
747 }
748
749 list_for_each_entry_safe(vif, tmp, &notify, notify_list) {
750 notify_remote_via_irq(vif->irq);
751 list_del_init(&vif->notify_list);
752 }
753
754 /* More work to do? */
755 if (!skb_queue_empty(&netbk->rx_queue) &&
756 !timer_pending(&netbk->net_timer))
757 xen_netbk_kick_thread(netbk);
758 }
759
760 void xen_netbk_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
761 {
762 struct xen_netbk *netbk = vif->netbk;
763
764 skb_queue_tail(&netbk->rx_queue, skb);
765
766 xen_netbk_kick_thread(netbk);
767 }
768
769 static void xen_netbk_alarm(unsigned long data)
770 {
771 struct xen_netbk *netbk = (struct xen_netbk *)data;
772 xen_netbk_kick_thread(netbk);
773 }
774
775 static int __on_net_schedule_list(struct xenvif *vif)
776 {
777 return !list_empty(&vif->schedule_list);
778 }
779
780 /* Must be called with net_schedule_list_lock held */
781 static void remove_from_net_schedule_list(struct xenvif *vif)
782 {
783 if (likely(__on_net_schedule_list(vif))) {
784 list_del_init(&vif->schedule_list);
785 xenvif_put(vif);
786 }
787 }
788
789 static struct xenvif *poll_net_schedule_list(struct xen_netbk *netbk)
790 {
791 struct xenvif *vif = NULL;
792
793 spin_lock_irq(&netbk->net_schedule_list_lock);
794 if (list_empty(&netbk->net_schedule_list))
795 goto out;
796
797 vif = list_first_entry(&netbk->net_schedule_list,
798 struct xenvif, schedule_list);
799 if (!vif)
800 goto out;
801
802 xenvif_get(vif);
803
804 remove_from_net_schedule_list(vif);
805 out:
806 spin_unlock_irq(&netbk->net_schedule_list_lock);
807 return vif;
808 }
809
810 void xen_netbk_schedule_xenvif(struct xenvif *vif)
811 {
812 unsigned long flags;
813 struct xen_netbk *netbk = vif->netbk;
814
815 if (__on_net_schedule_list(vif))
816 goto kick;
817
818 spin_lock_irqsave(&netbk->net_schedule_list_lock, flags);
819 if (!__on_net_schedule_list(vif) &&
820 likely(xenvif_schedulable(vif))) {
821 list_add_tail(&vif->schedule_list, &netbk->net_schedule_list);
822 xenvif_get(vif);
823 }
824 spin_unlock_irqrestore(&netbk->net_schedule_list_lock, flags);
825
826 kick:
827 smp_mb();
828 if ((nr_pending_reqs(netbk) < (MAX_PENDING_REQS/2)) &&
829 !list_empty(&netbk->net_schedule_list))
830 xen_netbk_kick_thread(netbk);
831 }
832
833 void xen_netbk_deschedule_xenvif(struct xenvif *vif)
834 {
835 struct xen_netbk *netbk = vif->netbk;
836 spin_lock_irq(&netbk->net_schedule_list_lock);
837 remove_from_net_schedule_list(vif);
838 spin_unlock_irq(&netbk->net_schedule_list_lock);
839 }
840
841 void xen_netbk_check_rx_xenvif(struct xenvif *vif)
842 {
843 int more_to_do;
844
845 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
846
847 if (more_to_do)
848 xen_netbk_schedule_xenvif(vif);
849 }
850
851 static void tx_add_credit(struct xenvif *vif)
852 {
853 unsigned long max_burst, max_credit;
854
855 /*
856 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
857 * Otherwise the interface can seize up due to insufficient credit.
858 */
859 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
860 max_burst = min(max_burst, 131072UL);
861 max_burst = max(max_burst, vif->credit_bytes);
862
863 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
864 max_credit = vif->remaining_credit + vif->credit_bytes;
865 if (max_credit < vif->remaining_credit)
866 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
867
868 vif->remaining_credit = min(max_credit, max_burst);
869 }
870
871 static void tx_credit_callback(unsigned long data)
872 {
873 struct xenvif *vif = (struct xenvif *)data;
874 tx_add_credit(vif);
875 xen_netbk_check_rx_xenvif(vif);
876 }
877
878 static void netbk_tx_err(struct xenvif *vif,
879 struct xen_netif_tx_request *txp, RING_IDX end)
880 {
881 RING_IDX cons = vif->tx.req_cons;
882
883 do {
884 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
885 if (cons == end)
886 break;
887 txp = RING_GET_REQUEST(&vif->tx, cons++);
888 } while (1);
889 vif->tx.req_cons = cons;
890 xen_netbk_check_rx_xenvif(vif);
891 xenvif_put(vif);
892 }
893
894 static void netbk_fatal_tx_err(struct xenvif *vif)
895 {
896 netdev_err(vif->dev, "fatal error; disabling device\n");
897 xenvif_carrier_off(vif);
898 xenvif_put(vif);
899 }
900
901 static int netbk_count_requests(struct xenvif *vif,
902 struct xen_netif_tx_request *first,
903 RING_IDX first_idx,
904 struct xen_netif_tx_request *txp,
905 int work_to_do)
906 {
907 RING_IDX cons = vif->tx.req_cons;
908 int slots = 0;
909 int drop_err = 0;
910
911 if (!(first->flags & XEN_NETTXF_more_data))
912 return 0;
913
914 do {
915 if (slots >= work_to_do) {
916 netdev_err(vif->dev,
917 "Asked for %d slots but exceeds this limit\n",
918 work_to_do);
919 netbk_fatal_tx_err(vif);
920 return -ENODATA;
921 }
922
923 /* This guest is really using too many slots and
924 * considered malicious.
925 */
926 if (unlikely(slots >= max_skb_slots)) {
927 netdev_err(vif->dev,
928 "Malicious frontend using %d slots, threshold %u\n",
929 slots, max_skb_slots);
930 netbk_fatal_tx_err(vif);
931 return -E2BIG;
932 }
933
934 /* Xen network protocol had implicit dependency on
935 * MAX_SKB_FRAGS. XEN_NETIF_NR_SLOTS_MIN is set to the
936 * historical MAX_SKB_FRAGS value 18 to honor the same
937 * behavior as before. Any packet using more than 18
938 * slots but less than max_skb_slots slots is dropped
939 */
940 if (!drop_err && slots >= XEN_NETIF_NR_SLOTS_MIN) {
941 if (net_ratelimit())
942 netdev_dbg(vif->dev,
943 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
944 slots, XEN_NETIF_NR_SLOTS_MIN);
945 drop_err = -E2BIG;
946 }
947
948 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
949 sizeof(*txp));
950
951 /* If the guest submitted a frame >= 64 KiB then
952 * first->size overflowed and following slots will
953 * appear to be larger than the frame.
954 *
955 * This cannot be fatal error as there are buggy
956 * frontends that do this.
957 *
958 * Consume all slots and drop the packet.
959 */
960 if (!drop_err && txp->size > first->size) {
961 if (net_ratelimit())
962 netdev_dbg(vif->dev,
963 "Invalid tx request, slot size %u > remaining size %u\n",
964 txp->size, first->size);
965 drop_err = -EIO;
966 }
967
968 first->size -= txp->size;
969 slots++;
970
971 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
972 netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
973 txp->offset, txp->size);
974 netbk_fatal_tx_err(vif);
975 return -EINVAL;
976 }
977 } while ((txp++)->flags & XEN_NETTXF_more_data);
978
979 if (drop_err) {
980 netbk_tx_err(vif, first, first_idx + slots);
981 return drop_err;
982 }
983
984 return slots;
985 }
986
987 static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
988 u16 pending_idx)
989 {
990 struct page *page;
991 page = alloc_page(GFP_KERNEL|__GFP_COLD);
992 if (!page)
993 return NULL;
994 set_page_ext(page, netbk, pending_idx);
995 netbk->mmap_pages[pending_idx] = page;
996 return page;
997 }
998
999 static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
1000 struct xenvif *vif,
1001 struct sk_buff *skb,
1002 struct xen_netif_tx_request *txp,
1003 struct gnttab_copy *gop)
1004 {
1005 struct skb_shared_info *shinfo = skb_shinfo(skb);
1006 skb_frag_t *frags = shinfo->frags;
1007 u16 pending_idx = *((u16 *)skb->data);
1008 u16 head_idx = 0;
1009 int slot, start;
1010 struct page *page;
1011 pending_ring_idx_t index, start_idx = 0;
1012 uint16_t dst_offset;
1013 unsigned int nr_slots;
1014 struct pending_tx_info *first = NULL;
1015
1016 /* At this point shinfo->nr_frags is in fact the number of
1017 * slots, which can be as large as XEN_NETIF_NR_SLOTS_MIN.
1018 */
1019 nr_slots = shinfo->nr_frags;
1020
1021 /* Skip first skb fragment if it is on same page as header fragment. */
1022 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
1023
1024 /* Coalesce tx requests, at this point the packet passed in
1025 * should be <= 64K. Any packets larger than 64K have been
1026 * handled in netbk_count_requests().
1027 */
1028 for (shinfo->nr_frags = slot = start; slot < nr_slots;
1029 shinfo->nr_frags++) {
1030 struct pending_tx_info *pending_tx_info =
1031 netbk->pending_tx_info;
1032
1033 page = alloc_page(GFP_KERNEL|__GFP_COLD);
1034 if (!page)
1035 goto err;
1036
1037 dst_offset = 0;
1038 first = NULL;
1039 while (dst_offset < PAGE_SIZE && slot < nr_slots) {
1040 gop->flags = GNTCOPY_source_gref;
1041
1042 gop->source.u.ref = txp->gref;
1043 gop->source.domid = vif->domid;
1044 gop->source.offset = txp->offset;
1045
1046 gop->dest.domid = DOMID_SELF;
1047
1048 gop->dest.offset = dst_offset;
1049 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1050
1051 if (dst_offset + txp->size > PAGE_SIZE) {
1052 /* This page can only merge a portion
1053 * of tx request. Do not increment any
1054 * pointer / counter here. The txp
1055 * will be dealt with in future
1056 * rounds, eventually hitting the
1057 * `else` branch.
1058 */
1059 gop->len = PAGE_SIZE - dst_offset;
1060 txp->offset += gop->len;
1061 txp->size -= gop->len;
1062 dst_offset += gop->len; /* quit loop */
1063 } else {
1064 /* This tx request can be merged in the page */
1065 gop->len = txp->size;
1066 dst_offset += gop->len;
1067
1068 index = pending_index(netbk->pending_cons++);
1069
1070 pending_idx = netbk->pending_ring[index];
1071
1072 memcpy(&pending_tx_info[pending_idx].req, txp,
1073 sizeof(*txp));
1074 xenvif_get(vif);
1075
1076 pending_tx_info[pending_idx].vif = vif;
1077
1078 /* Poison these fields, corresponding
1079 * fields for head tx req will be set
1080 * to correct values after the loop.
1081 */
1082 netbk->mmap_pages[pending_idx] = (void *)(~0UL);
1083 pending_tx_info[pending_idx].head =
1084 INVALID_PENDING_RING_IDX;
1085
1086 if (!first) {
1087 first = &pending_tx_info[pending_idx];
1088 start_idx = index;
1089 head_idx = pending_idx;
1090 }
1091
1092 txp++;
1093 slot++;
1094 }
1095
1096 gop++;
1097 }
1098
1099 first->req.offset = 0;
1100 first->req.size = dst_offset;
1101 first->head = start_idx;
1102 set_page_ext(page, netbk, head_idx);
1103 netbk->mmap_pages[head_idx] = page;
1104 frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
1105 }
1106
1107 BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
1108
1109 return gop;
1110 err:
1111 /* Unwind, freeing all pages and sending error responses. */
1112 while (shinfo->nr_frags-- > start) {
1113 xen_netbk_idx_release(netbk,
1114 frag_get_pending_idx(&frags[shinfo->nr_frags]),
1115 XEN_NETIF_RSP_ERROR);
1116 }
1117 /* The head too, if necessary. */
1118 if (start)
1119 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1120
1121 return NULL;
1122 }
1123
1124 static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
1125 struct sk_buff *skb,
1126 struct gnttab_copy **gopp)
1127 {
1128 struct gnttab_copy *gop = *gopp;
1129 u16 pending_idx = *((u16 *)skb->data);
1130 struct skb_shared_info *shinfo = skb_shinfo(skb);
1131 struct pending_tx_info *tx_info;
1132 int nr_frags = shinfo->nr_frags;
1133 int i, err, start;
1134 u16 peek; /* peek into next tx request */
1135
1136 /* Check status of header. */
1137 err = gop->status;
1138 if (unlikely(err))
1139 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1140
1141 /* Skip first skb fragment if it is on same page as header fragment. */
1142 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
1143
1144 for (i = start; i < nr_frags; i++) {
1145 int j, newerr;
1146 pending_ring_idx_t head;
1147
1148 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
1149 tx_info = &netbk->pending_tx_info[pending_idx];
1150 head = tx_info->head;
1151
1152 /* Check error status: if okay then remember grant handle. */
1153 do {
1154 newerr = (++gop)->status;
1155 if (newerr)
1156 break;
1157 peek = netbk->pending_ring[pending_index(++head)];
1158 } while (!pending_tx_is_head(netbk, peek));
1159
1160 if (likely(!newerr)) {
1161 /* Had a previous error? Invalidate this fragment. */
1162 if (unlikely(err))
1163 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1164 continue;
1165 }
1166
1167 /* Error on this fragment: respond to client with an error. */
1168 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1169
1170 /* Not the first error? Preceding frags already invalidated. */
1171 if (err)
1172 continue;
1173
1174 /* First error: invalidate header and preceding fragments. */
1175 pending_idx = *((u16 *)skb->data);
1176 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1177 for (j = start; j < i; j++) {
1178 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1179 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1180 }
1181
1182 /* Remember the error: invalidate all subsequent fragments. */
1183 err = newerr;
1184 }
1185
1186 *gopp = gop + 1;
1187 return err;
1188 }
1189
1190 static void xen_netbk_fill_frags(struct xen_netbk *netbk, struct sk_buff *skb)
1191 {
1192 struct skb_shared_info *shinfo = skb_shinfo(skb);
1193 int nr_frags = shinfo->nr_frags;
1194 int i;
1195
1196 for (i = 0; i < nr_frags; i++) {
1197 skb_frag_t *frag = shinfo->frags + i;
1198 struct xen_netif_tx_request *txp;
1199 struct page *page;
1200 u16 pending_idx;
1201
1202 pending_idx = frag_get_pending_idx(frag);
1203
1204 txp = &netbk->pending_tx_info[pending_idx].req;
1205 page = virt_to_page(idx_to_kaddr(netbk, pending_idx));
1206 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1207 skb->len += txp->size;
1208 skb->data_len += txp->size;
1209 skb->truesize += txp->size;
1210
1211 /* Take an extra reference to offset xen_netbk_idx_release */
1212 get_page(netbk->mmap_pages[pending_idx]);
1213 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1214 }
1215 }
1216
1217 static int xen_netbk_get_extras(struct xenvif *vif,
1218 struct xen_netif_extra_info *extras,
1219 int work_to_do)
1220 {
1221 struct xen_netif_extra_info extra;
1222 RING_IDX cons = vif->tx.req_cons;
1223
1224 do {
1225 if (unlikely(work_to_do-- <= 0)) {
1226 netdev_err(vif->dev, "Missing extra info\n");
1227 netbk_fatal_tx_err(vif);
1228 return -EBADR;
1229 }
1230
1231 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1232 sizeof(extra));
1233 if (unlikely(!extra.type ||
1234 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1235 vif->tx.req_cons = ++cons;
1236 netdev_err(vif->dev,
1237 "Invalid extra type: %d\n", extra.type);
1238 netbk_fatal_tx_err(vif);
1239 return -EINVAL;
1240 }
1241
1242 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1243 vif->tx.req_cons = ++cons;
1244 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1245
1246 return work_to_do;
1247 }
1248
1249 static int netbk_set_skb_gso(struct xenvif *vif,
1250 struct sk_buff *skb,
1251 struct xen_netif_extra_info *gso)
1252 {
1253 if (!gso->u.gso.size) {
1254 netdev_err(vif->dev, "GSO size must not be zero.\n");
1255 netbk_fatal_tx_err(vif);
1256 return -EINVAL;
1257 }
1258
1259 /* Currently only TCPv4 S.O. is supported. */
1260 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
1261 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1262 netbk_fatal_tx_err(vif);
1263 return -EINVAL;
1264 }
1265
1266 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1267 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1268
1269 /* Header must be checked, and gso_segs computed. */
1270 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1271 skb_shinfo(skb)->gso_segs = 0;
1272
1273 return 0;
1274 }
1275
1276 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1277 {
1278 struct iphdr *iph;
1279 unsigned char *th;
1280 int err = -EPROTO;
1281 int recalculate_partial_csum = 0;
1282
1283 /*
1284 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1285 * peers can fail to set NETRXF_csum_blank when sending a GSO
1286 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1287 * recalculate the partial checksum.
1288 */
1289 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1290 vif->rx_gso_checksum_fixup++;
1291 skb->ip_summed = CHECKSUM_PARTIAL;
1292 recalculate_partial_csum = 1;
1293 }
1294
1295 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1296 if (skb->ip_summed != CHECKSUM_PARTIAL)
1297 return 0;
1298
1299 if (skb->protocol != htons(ETH_P_IP))
1300 goto out;
1301
1302 iph = (void *)skb->data;
1303 th = skb->data + 4 * iph->ihl;
1304 if (th >= skb_tail_pointer(skb))
1305 goto out;
1306
1307 skb->csum_start = th - skb->head;
1308 switch (iph->protocol) {
1309 case IPPROTO_TCP:
1310 skb->csum_offset = offsetof(struct tcphdr, check);
1311
1312 if (recalculate_partial_csum) {
1313 struct tcphdr *tcph = (struct tcphdr *)th;
1314 tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1315 skb->len - iph->ihl*4,
1316 IPPROTO_TCP, 0);
1317 }
1318 break;
1319 case IPPROTO_UDP:
1320 skb->csum_offset = offsetof(struct udphdr, check);
1321
1322 if (recalculate_partial_csum) {
1323 struct udphdr *udph = (struct udphdr *)th;
1324 udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1325 skb->len - iph->ihl*4,
1326 IPPROTO_UDP, 0);
1327 }
1328 break;
1329 default:
1330 if (net_ratelimit())
1331 netdev_err(vif->dev,
1332 "Attempting to checksum a non-TCP/UDP packet, dropping a protocol %d packet\n",
1333 iph->protocol);
1334 goto out;
1335 }
1336
1337 if ((th + skb->csum_offset + 2) > skb_tail_pointer(skb))
1338 goto out;
1339
1340 err = 0;
1341
1342 out:
1343 return err;
1344 }
1345
1346 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1347 {
1348 u64 now = get_jiffies_64();
1349 u64 next_credit = vif->credit_window_start +
1350 msecs_to_jiffies(vif->credit_usec / 1000);
1351
1352 /* Timer could already be pending in rare cases. */
1353 if (timer_pending(&vif->credit_timeout))
1354 return true;
1355
1356 /* Passed the point where we can replenish credit? */
1357 if (time_after_eq64(now, next_credit)) {
1358 vif->credit_window_start = now;
1359 tx_add_credit(vif);
1360 }
1361
1362 /* Still too big to send right now? Set a callback. */
1363 if (size > vif->remaining_credit) {
1364 vif->credit_timeout.data =
1365 (unsigned long)vif;
1366 vif->credit_timeout.function =
1367 tx_credit_callback;
1368 mod_timer(&vif->credit_timeout,
1369 next_credit);
1370 vif->credit_window_start = next_credit;
1371
1372 return true;
1373 }
1374
1375 return false;
1376 }
1377
1378 static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
1379 {
1380 struct gnttab_copy *gop = netbk->tx_copy_ops, *request_gop;
1381 struct sk_buff *skb;
1382 int ret;
1383
1384 while ((nr_pending_reqs(netbk) + XEN_NETIF_NR_SLOTS_MIN
1385 < MAX_PENDING_REQS) &&
1386 !list_empty(&netbk->net_schedule_list)) {
1387 struct xenvif *vif;
1388 struct xen_netif_tx_request txreq;
1389 struct xen_netif_tx_request txfrags[max_skb_slots];
1390 struct page *page;
1391 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1392 u16 pending_idx;
1393 RING_IDX idx;
1394 int work_to_do;
1395 unsigned int data_len;
1396 pending_ring_idx_t index;
1397
1398 /* Get a netif from the list with work to do. */
1399 vif = poll_net_schedule_list(netbk);
1400 /* This can sometimes happen because the test of
1401 * list_empty(net_schedule_list) at the top of the
1402 * loop is unlocked. Just go back and have another
1403 * look.
1404 */
1405 if (!vif)
1406 continue;
1407
1408 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1409 XEN_NETIF_TX_RING_SIZE) {
1410 netdev_err(vif->dev,
1411 "Impossible number of requests. "
1412 "req_prod %d, req_cons %d, size %ld\n",
1413 vif->tx.sring->req_prod, vif->tx.req_cons,
1414 XEN_NETIF_TX_RING_SIZE);
1415 netbk_fatal_tx_err(vif);
1416 continue;
1417 }
1418
1419 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, work_to_do);
1420 if (!work_to_do) {
1421 xenvif_put(vif);
1422 continue;
1423 }
1424
1425 idx = vif->tx.req_cons;
1426 rmb(); /* Ensure that we see the request before we copy it. */
1427 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1428
1429 /* Credit-based scheduling. */
1430 if (txreq.size > vif->remaining_credit &&
1431 tx_credit_exceeded(vif, txreq.size)) {
1432 xenvif_put(vif);
1433 continue;
1434 }
1435
1436 vif->remaining_credit -= txreq.size;
1437
1438 work_to_do--;
1439 vif->tx.req_cons = ++idx;
1440
1441 memset(extras, 0, sizeof(extras));
1442 if (txreq.flags & XEN_NETTXF_extra_info) {
1443 work_to_do = xen_netbk_get_extras(vif, extras,
1444 work_to_do);
1445 idx = vif->tx.req_cons;
1446 if (unlikely(work_to_do < 0))
1447 continue;
1448 }
1449
1450 ret = netbk_count_requests(vif, &txreq, idx,
1451 txfrags, work_to_do);
1452 if (unlikely(ret < 0))
1453 continue;
1454
1455 idx += ret;
1456
1457 if (unlikely(txreq.size < ETH_HLEN)) {
1458 netdev_dbg(vif->dev,
1459 "Bad packet size: %d\n", txreq.size);
1460 netbk_tx_err(vif, &txreq, idx);
1461 continue;
1462 }
1463
1464 /* No crossing a page as the payload mustn't fragment. */
1465 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1466 netdev_err(vif->dev,
1467 "txreq.offset: %x, size: %u, end: %lu\n",
1468 txreq.offset, txreq.size,
1469 (txreq.offset&~PAGE_MASK) + txreq.size);
1470 netbk_fatal_tx_err(vif);
1471 continue;
1472 }
1473
1474 index = pending_index(netbk->pending_cons);
1475 pending_idx = netbk->pending_ring[index];
1476
1477 data_len = (txreq.size > PKT_PROT_LEN &&
1478 ret < XEN_NETIF_NR_SLOTS_MIN) ?
1479 PKT_PROT_LEN : txreq.size;
1480
1481 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1482 GFP_ATOMIC | __GFP_NOWARN);
1483 if (unlikely(skb == NULL)) {
1484 netdev_dbg(vif->dev,
1485 "Can't allocate a skb in start_xmit.\n");
1486 netbk_tx_err(vif, &txreq, idx);
1487 break;
1488 }
1489
1490 /* Packets passed to netif_rx() must have some headroom. */
1491 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1492
1493 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1494 struct xen_netif_extra_info *gso;
1495 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1496
1497 if (netbk_set_skb_gso(vif, skb, gso)) {
1498 /* Failure in netbk_set_skb_gso is fatal. */
1499 kfree_skb(skb);
1500 continue;
1501 }
1502 }
1503
1504 /* XXX could copy straight to head */
1505 page = xen_netbk_alloc_page(netbk, pending_idx);
1506 if (!page) {
1507 kfree_skb(skb);
1508 netbk_tx_err(vif, &txreq, idx);
1509 continue;
1510 }
1511
1512 gop->source.u.ref = txreq.gref;
1513 gop->source.domid = vif->domid;
1514 gop->source.offset = txreq.offset;
1515
1516 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1517 gop->dest.domid = DOMID_SELF;
1518 gop->dest.offset = txreq.offset;
1519
1520 gop->len = txreq.size;
1521 gop->flags = GNTCOPY_source_gref;
1522
1523 gop++;
1524
1525 memcpy(&netbk->pending_tx_info[pending_idx].req,
1526 &txreq, sizeof(txreq));
1527 netbk->pending_tx_info[pending_idx].vif = vif;
1528 netbk->pending_tx_info[pending_idx].head = index;
1529 *((u16 *)skb->data) = pending_idx;
1530
1531 __skb_put(skb, data_len);
1532
1533 skb_shinfo(skb)->nr_frags = ret;
1534 if (data_len < txreq.size) {
1535 skb_shinfo(skb)->nr_frags++;
1536 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1537 pending_idx);
1538 } else {
1539 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1540 INVALID_PENDING_IDX);
1541 }
1542
1543 __skb_queue_tail(&netbk->tx_queue, skb);
1544
1545 netbk->pending_cons++;
1546
1547 request_gop = xen_netbk_get_requests(netbk, vif,
1548 skb, txfrags, gop);
1549 if (request_gop == NULL) {
1550 kfree_skb(skb);
1551 netbk_tx_err(vif, &txreq, idx);
1552 continue;
1553 }
1554 gop = request_gop;
1555
1556 vif->tx.req_cons = idx;
1557 xen_netbk_check_rx_xenvif(vif);
1558
1559 if ((gop-netbk->tx_copy_ops) >= ARRAY_SIZE(netbk->tx_copy_ops))
1560 break;
1561 }
1562
1563 return gop - netbk->tx_copy_ops;
1564 }
1565
1566 static void xen_netbk_tx_submit(struct xen_netbk *netbk)
1567 {
1568 struct gnttab_copy *gop = netbk->tx_copy_ops;
1569 struct sk_buff *skb;
1570
1571 while ((skb = __skb_dequeue(&netbk->tx_queue)) != NULL) {
1572 struct xen_netif_tx_request *txp;
1573 struct xenvif *vif;
1574 u16 pending_idx;
1575 unsigned data_len;
1576
1577 pending_idx = *((u16 *)skb->data);
1578 vif = netbk->pending_tx_info[pending_idx].vif;
1579 txp = &netbk->pending_tx_info[pending_idx].req;
1580
1581 /* Check the remap error code. */
1582 if (unlikely(xen_netbk_tx_check_gop(netbk, skb, &gop))) {
1583 netdev_dbg(vif->dev, "netback grant failed.\n");
1584 skb_shinfo(skb)->nr_frags = 0;
1585 kfree_skb(skb);
1586 continue;
1587 }
1588
1589 data_len = skb->len;
1590 memcpy(skb->data,
1591 (void *)(idx_to_kaddr(netbk, pending_idx)|txp->offset),
1592 data_len);
1593 if (data_len < txp->size) {
1594 /* Append the packet payload as a fragment. */
1595 txp->offset += data_len;
1596 txp->size -= data_len;
1597 } else {
1598 /* Schedule a response immediately. */
1599 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
1600 }
1601
1602 if (txp->flags & XEN_NETTXF_csum_blank)
1603 skb->ip_summed = CHECKSUM_PARTIAL;
1604 else if (txp->flags & XEN_NETTXF_data_validated)
1605 skb->ip_summed = CHECKSUM_UNNECESSARY;
1606
1607 xen_netbk_fill_frags(netbk, skb);
1608
1609 /*
1610 * If the initial fragment was < PKT_PROT_LEN then
1611 * pull through some bytes from the other fragments to
1612 * increase the linear region to PKT_PROT_LEN bytes.
1613 */
1614 if (skb_headlen(skb) < PKT_PROT_LEN && skb_is_nonlinear(skb)) {
1615 int target = min_t(int, skb->len, PKT_PROT_LEN);
1616 __pskb_pull_tail(skb, target - skb_headlen(skb));
1617 }
1618
1619 skb->dev = vif->dev;
1620 skb->protocol = eth_type_trans(skb, skb->dev);
1621
1622 if (checksum_setup(vif, skb)) {
1623 netdev_dbg(vif->dev,
1624 "Can't setup checksum in net_tx_action\n");
1625 kfree_skb(skb);
1626 continue;
1627 }
1628
1629 vif->dev->stats.rx_bytes += skb->len;
1630 vif->dev->stats.rx_packets++;
1631
1632 xenvif_receive_skb(vif, skb);
1633 }
1634 }
1635
1636 /* Called after netfront has transmitted */
1637 static void xen_netbk_tx_action(struct xen_netbk *netbk)
1638 {
1639 unsigned nr_gops;
1640 int ret;
1641
1642 nr_gops = xen_netbk_tx_build_gops(netbk);
1643
1644 if (nr_gops == 0)
1645 return;
1646 ret = HYPERVISOR_grant_table_op(GNTTABOP_copy,
1647 netbk->tx_copy_ops, nr_gops);
1648 BUG_ON(ret);
1649
1650 xen_netbk_tx_submit(netbk);
1651
1652 }
1653
1654 static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
1655 u8 status)
1656 {
1657 struct xenvif *vif;
1658 struct pending_tx_info *pending_tx_info;
1659 pending_ring_idx_t head;
1660 u16 peek; /* peek into next tx request */
1661
1662 BUG_ON(netbk->mmap_pages[pending_idx] == (void *)(~0UL));
1663
1664 /* Already complete? */
1665 if (netbk->mmap_pages[pending_idx] == NULL)
1666 return;
1667
1668 pending_tx_info = &netbk->pending_tx_info[pending_idx];
1669
1670 vif = pending_tx_info->vif;
1671 head = pending_tx_info->head;
1672
1673 BUG_ON(!pending_tx_is_head(netbk, head));
1674 BUG_ON(netbk->pending_ring[pending_index(head)] != pending_idx);
1675
1676 do {
1677 pending_ring_idx_t index;
1678 pending_ring_idx_t idx = pending_index(head);
1679 u16 info_idx = netbk->pending_ring[idx];
1680
1681 pending_tx_info = &netbk->pending_tx_info[info_idx];
1682 make_tx_response(vif, &pending_tx_info->req, status);
1683
1684 /* Setting any number other than
1685 * INVALID_PENDING_RING_IDX indicates this slot is
1686 * starting a new packet / ending a previous packet.
1687 */
1688 pending_tx_info->head = 0;
1689
1690 index = pending_index(netbk->pending_prod++);
1691 netbk->pending_ring[index] = netbk->pending_ring[info_idx];
1692
1693 xenvif_put(vif);
1694
1695 peek = netbk->pending_ring[pending_index(++head)];
1696
1697 } while (!pending_tx_is_head(netbk, peek));
1698
1699 netbk->mmap_pages[pending_idx]->mapping = 0;
1700 put_page(netbk->mmap_pages[pending_idx]);
1701 netbk->mmap_pages[pending_idx] = NULL;
1702 }
1703
1704
1705 static void make_tx_response(struct xenvif *vif,
1706 struct xen_netif_tx_request *txp,
1707 s8 st)
1708 {
1709 RING_IDX i = vif->tx.rsp_prod_pvt;
1710 struct xen_netif_tx_response *resp;
1711 int notify;
1712
1713 resp = RING_GET_RESPONSE(&vif->tx, i);
1714 resp->id = txp->id;
1715 resp->status = st;
1716
1717 if (txp->flags & XEN_NETTXF_extra_info)
1718 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1719
1720 vif->tx.rsp_prod_pvt = ++i;
1721 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1722 if (notify)
1723 notify_remote_via_irq(vif->irq);
1724 }
1725
1726 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1727 u16 id,
1728 s8 st,
1729 u16 offset,
1730 u16 size,
1731 u16 flags)
1732 {
1733 RING_IDX i = vif->rx.rsp_prod_pvt;
1734 struct xen_netif_rx_response *resp;
1735
1736 resp = RING_GET_RESPONSE(&vif->rx, i);
1737 resp->offset = offset;
1738 resp->flags = flags;
1739 resp->id = id;
1740 resp->status = (s16)size;
1741 if (st < 0)
1742 resp->status = (s16)st;
1743
1744 vif->rx.rsp_prod_pvt = ++i;
1745
1746 return resp;
1747 }
1748
1749 static inline int rx_work_todo(struct xen_netbk *netbk)
1750 {
1751 return !skb_queue_empty(&netbk->rx_queue);
1752 }
1753
1754 static inline int tx_work_todo(struct xen_netbk *netbk)
1755 {
1756
1757 if ((nr_pending_reqs(netbk) + XEN_NETIF_NR_SLOTS_MIN
1758 < MAX_PENDING_REQS) &&
1759 !list_empty(&netbk->net_schedule_list))
1760 return 1;
1761
1762 return 0;
1763 }
1764
1765 static int xen_netbk_kthread(void *data)
1766 {
1767 struct xen_netbk *netbk = data;
1768 while (!kthread_should_stop()) {
1769 wait_event_interruptible(netbk->wq,
1770 rx_work_todo(netbk) ||
1771 tx_work_todo(netbk) ||
1772 kthread_should_stop());
1773 cond_resched();
1774
1775 if (kthread_should_stop())
1776 break;
1777
1778 if (rx_work_todo(netbk))
1779 xen_netbk_rx_action(netbk);
1780
1781 if (tx_work_todo(netbk))
1782 xen_netbk_tx_action(netbk);
1783 }
1784
1785 return 0;
1786 }
1787
1788 void xen_netbk_unmap_frontend_rings(struct xenvif *vif)
1789 {
1790 if (vif->tx.sring)
1791 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1792 vif->tx.sring);
1793 if (vif->rx.sring)
1794 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1795 vif->rx.sring);
1796 }
1797
1798 int xen_netbk_map_frontend_rings(struct xenvif *vif,
1799 grant_ref_t tx_ring_ref,
1800 grant_ref_t rx_ring_ref)
1801 {
1802 void *addr;
1803 struct xen_netif_tx_sring *txs;
1804 struct xen_netif_rx_sring *rxs;
1805
1806 int err = -ENOMEM;
1807
1808 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1809 tx_ring_ref, &addr);
1810 if (err)
1811 goto err;
1812
1813 txs = (struct xen_netif_tx_sring *)addr;
1814 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1815
1816 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1817 rx_ring_ref, &addr);
1818 if (err)
1819 goto err;
1820
1821 rxs = (struct xen_netif_rx_sring *)addr;
1822 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1823
1824 vif->rx_req_cons_peek = 0;
1825
1826 return 0;
1827
1828 err:
1829 xen_netbk_unmap_frontend_rings(vif);
1830 return err;
1831 }
1832
1833 static int __init netback_init(void)
1834 {
1835 int i;
1836 int rc = 0;
1837 int group;
1838
1839 if (!xen_domain())
1840 return -ENODEV;
1841
1842 if (max_skb_slots < XEN_NETIF_NR_SLOTS_MIN) {
1843 printk(KERN_INFO
1844 "xen-netback: max_skb_slots too small (%d), bump it to XEN_NETIF_NR_SLOTS_MIN (%d)\n",
1845 max_skb_slots, XEN_NETIF_NR_SLOTS_MIN);
1846 max_skb_slots = XEN_NETIF_NR_SLOTS_MIN;
1847 }
1848
1849 xen_netbk_group_nr = num_online_cpus();
1850 xen_netbk = vzalloc(sizeof(struct xen_netbk) * xen_netbk_group_nr);
1851 if (!xen_netbk)
1852 return -ENOMEM;
1853
1854 for (group = 0; group < xen_netbk_group_nr; group++) {
1855 struct xen_netbk *netbk = &xen_netbk[group];
1856 skb_queue_head_init(&netbk->rx_queue);
1857 skb_queue_head_init(&netbk->tx_queue);
1858
1859 init_timer(&netbk->net_timer);
1860 netbk->net_timer.data = (unsigned long)netbk;
1861 netbk->net_timer.function = xen_netbk_alarm;
1862
1863 netbk->pending_cons = 0;
1864 netbk->pending_prod = MAX_PENDING_REQS;
1865 for (i = 0; i < MAX_PENDING_REQS; i++)
1866 netbk->pending_ring[i] = i;
1867
1868 init_waitqueue_head(&netbk->wq);
1869 netbk->task = kthread_create(xen_netbk_kthread,
1870 (void *)netbk,
1871 "netback/%u", group);
1872
1873 if (IS_ERR(netbk->task)) {
1874 printk(KERN_ALERT "kthread_create() fails at netback\n");
1875 del_timer(&netbk->net_timer);
1876 rc = PTR_ERR(netbk->task);
1877 goto failed_init;
1878 }
1879
1880 kthread_bind(netbk->task, group);
1881
1882 INIT_LIST_HEAD(&netbk->net_schedule_list);
1883
1884 spin_lock_init(&netbk->net_schedule_list_lock);
1885
1886 atomic_set(&netbk->netfront_count, 0);
1887
1888 wake_up_process(netbk->task);
1889 }
1890
1891 rc = xenvif_xenbus_init();
1892 if (rc)
1893 goto failed_init;
1894
1895 return 0;
1896
1897 failed_init:
1898 while (--group >= 0) {
1899 struct xen_netbk *netbk = &xen_netbk[group];
1900 for (i = 0; i < MAX_PENDING_REQS; i++) {
1901 if (netbk->mmap_pages[i])
1902 __free_page(netbk->mmap_pages[i]);
1903 }
1904 del_timer(&netbk->net_timer);
1905 kthread_stop(netbk->task);
1906 }
1907 vfree(xen_netbk);
1908 return rc;
1909
1910 }
1911
1912 module_init(netback_init);
1913
1914 MODULE_LICENSE("Dual BSD/GPL");
1915 MODULE_ALIAS("xen-backend:vif");
Linux with added FPC-III support