projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
xen-netfront: reduce gso_max_size to account for max TCP header
[linux-2.6-block.git] / drivers / net / xen-netback / netback.c
CommitLineData
f942dc25
IC		 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
ca981633 43#include <xen/xen.h>
f942dc25
IC		 44#include <xen/events.h>
45#include <xen/interface/memory.h>
46
47#include <asm/xen/hypercall.h>
48#include <asm/xen/page.h>
49
50struct pending_tx_info {
51 struct xen_netif_tx_request req;
52 struct xenvif *vif;
53};
54typedef unsigned int pending_ring_idx_t;
55
56struct netbk_rx_meta {
57 int id;
58 int size;
59 int gso_size;
60};
61
62#define MAX_PENDING_REQS 256
63
ea066ad1
IC		 64/* Discriminate from any valid pending_idx value. */
65#define INVALID_PENDING_IDX 0xFFFF
66
f942dc25
IC		 67#define MAX_BUFFER_OFFSET PAGE_SIZE
68
69/* extra field used in struct page */
70union page_ext {
71 struct {
72#if BITS_PER_LONG < 64
73#define IDX_WIDTH 8
74#define GROUP_WIDTH (BITS_PER_LONG - IDX_WIDTH)
75 unsigned int group:GROUP_WIDTH;
76 unsigned int idx:IDX_WIDTH;
77#else
78 unsigned int group, idx;
79#endif
80 } e;
81 void *mapping;
82};
83
84struct xen_netbk {
85 wait_queue_head_t wq;
86 struct task_struct *task;
87
88 struct sk_buff_head rx_queue;
89 struct sk_buff_head tx_queue;
90
91 struct timer_list net_timer;
92
93 struct page *mmap_pages[MAX_PENDING_REQS];
94
95 pending_ring_idx_t pending_prod;
96 pending_ring_idx_t pending_cons;
97 struct list_head net_schedule_list;
98
99 /* Protect the net_schedule_list in netif. */
100 spinlock_t net_schedule_list_lock;
101
102 atomic_t netfront_count;
103
104 struct pending_tx_info pending_tx_info[MAX_PENDING_REQS];
105 struct gnttab_copy tx_copy_ops[MAX_PENDING_REQS];
106
107 u16 pending_ring[MAX_PENDING_REQS];
108
109 /*
110 * Given MAX_BUFFER_OFFSET of 4096 the worst case is that each
111 * head/fragment page uses 2 copy operations because it
112 * straddles two buffers in the frontend.
113 */
114 struct gnttab_copy grant_copy_op[2*XEN_NETIF_RX_RING_SIZE];
115 struct netbk_rx_meta meta[2*XEN_NETIF_RX_RING_SIZE];
116};
117
118static struct xen_netbk *xen_netbk;
119static int xen_netbk_group_nr;
120
121void xen_netbk_add_xenvif(struct xenvif *vif)
122{
123 int i;
124 int min_netfront_count;
125 int min_group = 0;
126 struct xen_netbk *netbk;
127
128 min_netfront_count = atomic_read(&xen_netbk[0].netfront_count);
129 for (i = 0; i < xen_netbk_group_nr; i++) {
130 int netfront_count = atomic_read(&xen_netbk[i].netfront_count);
131 if (netfront_count < min_netfront_count) {
132 min_group = i;
133 min_netfront_count = netfront_count;
134 }
135 }
136
137 netbk = &xen_netbk[min_group];
138
139 vif->netbk = netbk;
140 atomic_inc(&netbk->netfront_count);
141}
142
143void xen_netbk_remove_xenvif(struct xenvif *vif)
144{
145 struct xen_netbk *netbk = vif->netbk;
146 vif->netbk = NULL;
147 atomic_dec(&netbk->netfront_count);
148}
149
7d5145d8
MD		 150static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
151 u8 status);
f942dc25
IC		 152static void make_tx_response(struct xenvif *vif,
153 struct xen_netif_tx_request *txp,
154 s8 st);
155static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
156 u16 id,
157 s8 st,
158 u16 offset,
159 u16 size,
160 u16 flags);
161
162static inline unsigned long idx_to_pfn(struct xen_netbk *netbk,
ea066ad1 163 u16 idx)
f942dc25
IC		 164{
165 return page_to_pfn(netbk->mmap_pages[idx]);
166}
167
168static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
ea066ad1 169 u16 idx)
f942dc25
IC		 170{
171 return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
172}
173
174/* extra field used in struct page */
175static inline void set_page_ext(struct page *pg, struct xen_netbk *netbk,
176 unsigned int idx)
177{
178 unsigned int group = netbk - xen_netbk;
179 union page_ext ext = { .e = { .group = group + 1, .idx = idx } };
180
181 BUILD_BUG_ON(sizeof(ext) > sizeof(ext.mapping));
182 pg->mapping = ext.mapping;
183}
184
185static int get_page_ext(struct page *pg,
186 unsigned int *pgroup, unsigned int *pidx)
187{
188 union page_ext ext = { .mapping = pg->mapping };
189 struct xen_netbk *netbk;
190 unsigned int group, idx;
191
192 group = ext.e.group - 1;
193
194 if (group < 0 || group >= xen_netbk_group_nr)
195 return 0;
196
197 netbk = &xen_netbk[group];
198
199 idx = ext.e.idx;
200
201 if ((idx < 0) || (idx >= MAX_PENDING_REQS))
202 return 0;
203
204 if (netbk->mmap_pages[idx] != pg)
205 return 0;
206
207 *pgroup = group;
208 *pidx = idx;
209
210 return 1;
211}
212
213/*
214 * This is the amount of packet we copy rather than map, so that the
215 * guest can't fiddle with the contents of the headers while we do
216 * packet processing on them (netfilter, routing, etc).
217 */
218#define PKT_PROT_LEN (ETH_HLEN + \
219 VLAN_HLEN + \
220 sizeof(struct iphdr) + MAX_IPOPTLEN + \
221 sizeof(struct tcphdr) + MAX_TCP_OPTION_SPACE)
222
ea066ad1
IC		 223static u16 frag_get_pending_idx(skb_frag_t *frag)
224{
225 return (u16)frag->page_offset;
226}
227
228static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
229{
230 frag->page_offset = pending_idx;
231}
232
f942dc25
IC		 233static inline pending_ring_idx_t pending_index(unsigned i)
234{
235 return i & (MAX_PENDING_REQS-1);
236}
237
238static inline pending_ring_idx_t nr_pending_reqs(struct xen_netbk *netbk)
239{
240 return MAX_PENDING_REQS -
241 netbk->pending_prod + netbk->pending_cons;
242}
243
244static void xen_netbk_kick_thread(struct xen_netbk *netbk)
245{
246 wake_up(&netbk->wq);
247}
248
249static int max_required_rx_slots(struct xenvif *vif)
250{
251 int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
252
253 if (vif->can_sg || vif->gso || vif->gso_prefix)
254 max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
255
256 return max;
257}
258
259int xen_netbk_rx_ring_full(struct xenvif *vif)
260{
261 RING_IDX peek = vif->rx_req_cons_peek;
262 RING_IDX needed = max_required_rx_slots(vif);
263
264 return ((vif->rx.sring->req_prod - peek) < needed) ||
265 ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
266}
267
268int xen_netbk_must_stop_queue(struct xenvif *vif)
269{
270 if (!xen_netbk_rx_ring_full(vif))
271 return 0;
272
273 vif->rx.sring->req_event = vif->rx_req_cons_peek +
274 max_required_rx_slots(vif);
275 mb(); /* request notification /then/ check the queue */
276
277 return xen_netbk_rx_ring_full(vif);
278}
279
280/*
281 * Returns true if we should start a new receive buffer instead of
282 * adding 'size' bytes to a buffer which currently contains 'offset'
283 * bytes.
284 */
285static bool start_new_rx_buffer(int offset, unsigned long size, int head)
286{
287 /* simple case: we have completely filled the current buffer. */
288 if (offset == MAX_BUFFER_OFFSET)
289 return true;
290
291 /*
292 * complex case: start a fresh buffer if the current frag
293 * would overflow the current buffer but only if:
294 * (i) this frag would fit completely in the next buffer
295 * and (ii) there is already some data in the current buffer
296 * and (iii) this is not the head buffer.
297 *
298 * Where:
299 * - (i) stops us splitting a frag into two copies
300 * unless the frag is too large for a single buffer.
301 * - (ii) stops us from leaving a buffer pointlessly empty.
302 * - (iii) stops us leaving the first buffer
303 * empty. Strictly speaking this is already covered
304 * by (ii) but is explicitly checked because
305 * netfront relies on the first buffer being
306 * non-empty and can crash otherwise.
307 *
308 * This means we will effectively linearise small
309 * frags but do not needlessly split large buffers
310 * into multiple copies tend to give large frags their
311 * own buffers as before.
312 */
313 if ((offset + size > MAX_BUFFER_OFFSET) &&
314 (size <= MAX_BUFFER_OFFSET) && offset && !head)
315 return true;
316
317 return false;
318}
319
320/*
321 * Figure out how many ring slots we're going to need to send @skb to
322 * the guest. This function is essentially a dry run of
323 * netbk_gop_frag_copy.
324 */
325unsigned int xen_netbk_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
326{
327 unsigned int count;
328 int i, copy_off;
329
e26b203e 330 count = DIV_ROUND_UP(skb_headlen(skb), PAGE_SIZE);
f942dc25
IC		 331
332 copy_off = skb_headlen(skb) % PAGE_SIZE;
333
334 if (skb_shinfo(skb)->gso_size)
335 count++;
336
337 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
9e903e08 338 unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
6a8ed462 339 unsigned long offset = skb_shinfo(skb)->frags[i].page_offset;
f942dc25 340 unsigned long bytes;
6a8ed462
IC		 341
342 offset &= ~PAGE_MASK;
343
f942dc25 344 while (size > 0) {
6a8ed462 345 BUG_ON(offset >= PAGE_SIZE);
f942dc25
IC		 346 BUG_ON(copy_off > MAX_BUFFER_OFFSET);
347
6a8ed462
IC		 348 bytes = PAGE_SIZE - offset;
349
350 if (bytes > size)
351 bytes = size;
352
353 if (start_new_rx_buffer(copy_off, bytes, 0)) {
f942dc25
IC		 354 count++;
355 copy_off = 0;
356 }
357
f942dc25
IC		 358 if (copy_off + bytes > MAX_BUFFER_OFFSET)
359 bytes = MAX_BUFFER_OFFSET - copy_off;
360
361 copy_off += bytes;
6a8ed462
IC		 362
363 offset += bytes;
f942dc25 364 size -= bytes;
6a8ed462
IC		 365
366 if (offset == PAGE_SIZE)
367 offset = 0;
f942dc25
IC		 368 }
369 }
370 return count;
371}
372
373struct netrx_pending_operations {
374 unsigned copy_prod, copy_cons;
375 unsigned meta_prod, meta_cons;
376 struct gnttab_copy *copy;
377 struct netbk_rx_meta *meta;
378 int copy_off;
379 grant_ref_t copy_gref;
380};
381
382static struct netbk_rx_meta *get_next_rx_buffer(struct xenvif *vif,
383 struct netrx_pending_operations *npo)
384{
385 struct netbk_rx_meta *meta;
386 struct xen_netif_rx_request *req;
387
388 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
389
390 meta = npo->meta + npo->meta_prod++;
391 meta->gso_size = 0;
392 meta->size = 0;
393 meta->id = req->id;
394
395 npo->copy_off = 0;
396 npo->copy_gref = req->gref;
397
398 return meta;
399}
400
401/*
402 * Set up the grant operations for this fragment. If it's a flipping
403 * interface, we also set up the unmap request from here.
404 */
405static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
406 struct netrx_pending_operations *npo,
407 struct page *page, unsigned long size,
408 unsigned long offset, int *head)
409{
410 struct gnttab_copy *copy_gop;
411 struct netbk_rx_meta *meta;
412 /*
e34c0246 413 * These variables are used iff get_page_ext returns true,
f942dc25
IC		 414 * in which case they are guaranteed to be initialized.
415 */
416 unsigned int uninitialized_var(group), uninitialized_var(idx);
417 int foreign = get_page_ext(page, &group, &idx);
418 unsigned long bytes;
419
420 /* Data must not cross a page boundary. */
6a8ed462 421 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
f942dc25
IC		 422
423 meta = npo->meta + npo->meta_prod - 1;
424
6a8ed462
IC		 425 /* Skip unused frames from start of page */
426 page += offset >> PAGE_SHIFT;
427 offset &= ~PAGE_MASK;
428
f942dc25 429 while (size > 0) {
6a8ed462 430 BUG_ON(offset >= PAGE_SIZE);
f942dc25
IC		 431 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
432
6a8ed462
IC		 433 bytes = PAGE_SIZE - offset;
434
435 if (bytes > size)
436 bytes = size;
437
438 if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
f942dc25
IC		 439 /*
440 * Netfront requires there to be some data in the head
441 * buffer.
442 */
443 BUG_ON(*head);
444
445 meta = get_next_rx_buffer(vif, npo);
446 }
447
f942dc25
IC		 448 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
449 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
450
451 copy_gop = npo->copy + npo->copy_prod++;
452 copy_gop->flags = GNTCOPY_dest_gref;
453 if (foreign) {
454 struct xen_netbk *netbk = &xen_netbk[group];
455 struct pending_tx_info *src_pend;
456
457 src_pend = &netbk->pending_tx_info[idx];
458
459 copy_gop->source.domid = src_pend->vif->domid;
460 copy_gop->source.u.ref = src_pend->req.gref;
461 copy_gop->flags |= GNTCOPY_source_gref;
462 } else {
463 void *vaddr = page_address(page);
464 copy_gop->source.domid = DOMID_SELF;
465 copy_gop->source.u.gmfn = virt_to_mfn(vaddr);
466 }
467 copy_gop->source.offset = offset;
468 copy_gop->dest.domid = vif->domid;
469
470 copy_gop->dest.offset = npo->copy_off;
471 copy_gop->dest.u.ref = npo->copy_gref;
472 copy_gop->len = bytes;
473
474 npo->copy_off += bytes;
475 meta->size += bytes;
476
477 offset += bytes;
478 size -= bytes;
479
6a8ed462
IC		 480 /* Next frame */
481 if (offset == PAGE_SIZE && size) {
482 BUG_ON(!PageCompound(page));
483 page++;
484 offset = 0;
485 }
486
f942dc25
IC		 487 /* Leave a gap for the GSO descriptor. */
488 if (*head && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
489 vif->rx.req_cons++;
490
491 *head = 0; /* There must be something in this buffer now. */
492
493 }
494}
495
496/*
497 * Prepare an SKB to be transmitted to the frontend.
498 *
499 * This function is responsible for allocating grant operations, meta
500 * structures, etc.
501 *
502 * It returns the number of meta structures consumed. The number of
503 * ring slots used is always equal to the number of meta slots used
504 * plus the number of GSO descriptors used. Currently, we use either
505 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
506 * frontend-side LRO).
507 */
508static int netbk_gop_skb(struct sk_buff *skb,
509 struct netrx_pending_operations *npo)
510{
511 struct xenvif *vif = netdev_priv(skb->dev);
512 int nr_frags = skb_shinfo(skb)->nr_frags;
513 int i;
514 struct xen_netif_rx_request *req;
515 struct netbk_rx_meta *meta;
516 unsigned char *data;
517 int head = 1;
518 int old_meta_prod;
519
520 old_meta_prod = npo->meta_prod;
521
522 /* Set up a GSO prefix descriptor, if necessary */
523 if (skb_shinfo(skb)->gso_size && vif->gso_prefix) {
524 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
525 meta = npo->meta + npo->meta_prod++;
526 meta->gso_size = skb_shinfo(skb)->gso_size;
527 meta->size = 0;
528 meta->id = req->id;
529 }
530
531 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
532 meta = npo->meta + npo->meta_prod++;
533
534 if (!vif->gso_prefix)
535 meta->gso_size = skb_shinfo(skb)->gso_size;
536 else
537 meta->gso_size = 0;
538
539 meta->size = 0;
540 meta->id = req->id;
541 npo->copy_off = 0;
542 npo->copy_gref = req->gref;
543
544 data = skb->data;
545 while (data < skb_tail_pointer(skb)) {
546 unsigned int offset = offset_in_page(data);
547 unsigned int len = PAGE_SIZE - offset;
548
549 if (data + len > skb_tail_pointer(skb))
550 len = skb_tail_pointer(skb) - data;
551
552 netbk_gop_frag_copy(vif, skb, npo,
553 virt_to_page(data), len, offset, &head);
554 data += len;
555 }
556
557 for (i = 0; i < nr_frags; i++) {
558 netbk_gop_frag_copy(vif, skb, npo,
ea066ad1 559 skb_frag_page(&skb_shinfo(skb)->frags[i]),
9e903e08 560 skb_frag_size(&skb_shinfo(skb)->frags[i]),
f942dc25
IC		 561 skb_shinfo(skb)->frags[i].page_offset,
562 &head);
563 }
564
565 return npo->meta_prod - old_meta_prod;
566}
567
568/*
569 * This is a twin to netbk_gop_skb. Assume that netbk_gop_skb was
570 * used to set up the operations on the top of
571 * netrx_pending_operations, which have since been done. Check that
572 * they didn't give any errors and advance over them.
573 */
574static int netbk_check_gop(struct xenvif *vif, int nr_meta_slots,
575 struct netrx_pending_operations *npo)
576{
577 struct gnttab_copy *copy_op;
578 int status = XEN_NETIF_RSP_OKAY;
579 int i;
580
581 for (i = 0; i < nr_meta_slots; i++) {
582 copy_op = npo->copy + npo->copy_cons++;
583 if (copy_op->status != GNTST_okay) {
584 netdev_dbg(vif->dev,
585 "Bad status %d from copy to DOM%d.\n",
586 copy_op->status, vif->domid);
587 status = XEN_NETIF_RSP_ERROR;
588 }
589 }
590
591 return status;
592}
593
594static void netbk_add_frag_responses(struct xenvif *vif, int status,
595 struct netbk_rx_meta *meta,
596 int nr_meta_slots)
597{
598 int i;
599 unsigned long offset;
600
601 /* No fragments used */
602 if (nr_meta_slots <= 1)
603 return;
604
605 nr_meta_slots--;
606
607 for (i = 0; i < nr_meta_slots; i++) {
608 int flags;
609 if (i == nr_meta_slots - 1)
610 flags = 0;
611 else
612 flags = XEN_NETRXF_more_data;
613
614 offset = 0;
615 make_rx_response(vif, meta[i].id, status, offset,
616 meta[i].size, flags);
617 }
618}
619
620struct skb_cb_overlay {
621 int meta_slots_used;
622};
623
624static void xen_netbk_rx_action(struct xen_netbk *netbk)
625{
626 struct xenvif *vif = NULL, *tmp;
627 s8 status;
628 u16 irq, flags;
629 struct xen_netif_rx_response *resp;
630 struct sk_buff_head rxq;
631 struct sk_buff *skb;
632 LIST_HEAD(notify);
633 int ret;
634 int nr_frags;
635 int count;
636 unsigned long offset;
637 struct skb_cb_overlay *sco;
638
639 struct netrx_pending_operations npo = {
640 .copy = netbk->grant_copy_op,
641 .meta = netbk->meta,
642 };
643
644 skb_queue_head_init(&rxq);
645
646 count = 0;
647
648 while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
649 vif = netdev_priv(skb->dev);
650 nr_frags = skb_shinfo(skb)->nr_frags;
651
652 sco = (struct skb_cb_overlay *)skb->cb;
653 sco->meta_slots_used = netbk_gop_skb(skb, &npo);
654
655 count += nr_frags + 1;
656
657 __skb_queue_tail(&rxq, skb);
658
659 /* Filled the batch queue? */
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));
c571898f 670 gnttab_batch_copy(netbk->grant_copy_op, npo.copy_prod);
f942dc25
IC		 671
672 while ((skb = __skb_dequeue(&rxq)) != NULL) {
673 sco = (struct skb_cb_overlay *)skb->cb;
674
675 vif = netdev_priv(skb->dev);
676
677 if (netbk->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
678 resp = RING_GET_RESPONSE(&vif->rx,
679 vif->rx.rsp_prod_pvt++);
680
681 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
682
683 resp->offset = netbk->meta[npo.meta_cons].gso_size;
684 resp->id = netbk->meta[npo.meta_cons].id;
685 resp->status = sco->meta_slots_used;
686
687 npo.meta_cons++;
688 sco->meta_slots_used--;
689 }
690
691
692 vif->dev->stats.tx_bytes += skb->len;
693 vif->dev->stats.tx_packets++;
694
695 status = netbk_check_gop(vif, sco->meta_slots_used, &npo);
696
697 if (sco->meta_slots_used == 1)
698 flags = 0;
699 else
700 flags = XEN_NETRXF_more_data;
701
702 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
703 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
704 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
705 /* remote but checksummed. */
706 flags |= XEN_NETRXF_data_validated;
707
708 offset = 0;
709 resp = make_rx_response(vif, netbk->meta[npo.meta_cons].id,
710 status, offset,
711 netbk->meta[npo.meta_cons].size,
712 flags);
713
714 if (netbk->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
715 struct xen_netif_extra_info *gso =
716 (struct xen_netif_extra_info *)
717 RING_GET_RESPONSE(&vif->rx,
718 vif->rx.rsp_prod_pvt++);
719
720 resp->flags |= XEN_NETRXF_extra_info;
721
722 gso->u.gso.size = netbk->meta[npo.meta_cons].gso_size;
723 gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
724 gso->u.gso.pad = 0;
725 gso->u.gso.features = 0;
726
727 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
728 gso->flags = 0;
729 }
730
731 netbk_add_frag_responses(vif, status,
732 netbk->meta + npo.meta_cons + 1,
733 sco->meta_slots_used);
734
735 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
736 irq = vif->irq;
737 if (ret && list_empty(&vif->notify_list))
738 list_add_tail(&vif->notify_list, &notify);
739
740 xenvif_notify_tx_completion(vif);
741
742 xenvif_put(vif);
743 npo.meta_cons += sco->meta_slots_used;
744 dev_kfree_skb(skb);
745 }
746
747 list_for_each_entry_safe(vif, tmp, &notify, notify_list) {
748 notify_remote_via_irq(vif->irq);
749 list_del_init(&vif->notify_list);
750 }
751
752 /* More work to do? */
753 if (!skb_queue_empty(&netbk->rx_queue) &&
754 !timer_pending(&netbk->net_timer))
755 xen_netbk_kick_thread(netbk);
756}
757
758void xen_netbk_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
759{
760 struct xen_netbk *netbk = vif->netbk;
761
762 skb_queue_tail(&netbk->rx_queue, skb);
763
764 xen_netbk_kick_thread(netbk);
765}
766
767static void xen_netbk_alarm(unsigned long data)
768{
769 struct xen_netbk *netbk = (struct xen_netbk *)data;
770 xen_netbk_kick_thread(netbk);
771}
772
773static int __on_net_schedule_list(struct xenvif *vif)
774{
775 return !list_empty(&vif->schedule_list);
776}
777
778/* Must be called with net_schedule_list_lock held */
779static void remove_from_net_schedule_list(struct xenvif *vif)
780{
781 if (likely(__on_net_schedule_list(vif))) {
782 list_del_init(&vif->schedule_list);
783 xenvif_put(vif);
784 }
785}
786
787static struct xenvif *poll_net_schedule_list(struct xen_netbk *netbk)
788{
789 struct xenvif *vif = NULL;
790
791 spin_lock_irq(&netbk->net_schedule_list_lock);
792 if (list_empty(&netbk->net_schedule_list))
793 goto out;
794
795 vif = list_first_entry(&netbk->net_schedule_list,
796 struct xenvif, schedule_list);
797 if (!vif)
798 goto out;
799
800 xenvif_get(vif);
801
802 remove_from_net_schedule_list(vif);
803out:
804 spin_unlock_irq(&netbk->net_schedule_list_lock);
805 return vif;
806}
807
808void xen_netbk_schedule_xenvif(struct xenvif *vif)
809{
810 unsigned long flags;
811 struct xen_netbk *netbk = vif->netbk;
812
813 if (__on_net_schedule_list(vif))
814 goto kick;
815
816 spin_lock_irqsave(&netbk->net_schedule_list_lock, flags);
817 if (!__on_net_schedule_list(vif) &&
818 likely(xenvif_schedulable(vif))) {
819 list_add_tail(&vif->schedule_list, &netbk->net_schedule_list);
820 xenvif_get(vif);
821 }
822 spin_unlock_irqrestore(&netbk->net_schedule_list_lock, flags);
823
824kick:
825 smp_mb();
826 if ((nr_pending_reqs(netbk) < (MAX_PENDING_REQS/2)) &&
827 !list_empty(&netbk->net_schedule_list))
828 xen_netbk_kick_thread(netbk);
829}
830
831void xen_netbk_deschedule_xenvif(struct xenvif *vif)
832{
833 struct xen_netbk *netbk = vif->netbk;
834 spin_lock_irq(&netbk->net_schedule_list_lock);
835 remove_from_net_schedule_list(vif);
836 spin_unlock_irq(&netbk->net_schedule_list_lock);
837}
838
839void xen_netbk_check_rx_xenvif(struct xenvif *vif)
840{
841 int more_to_do;
842
843 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
844
845 if (more_to_do)
846 xen_netbk_schedule_xenvif(vif);
847}
848
849static void tx_add_credit(struct xenvif *vif)
850{
851 unsigned long max_burst, max_credit;
852
853 /*
854 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
855 * Otherwise the interface can seize up due to insufficient credit.
856 */
857 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
858 max_burst = min(max_burst, 131072UL);
859 max_burst = max(max_burst, vif->credit_bytes);
860
861 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
862 max_credit = vif->remaining_credit + vif->credit_bytes;
863 if (max_credit < vif->remaining_credit)
864 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
865
866 vif->remaining_credit = min(max_credit, max_burst);
867}
868
869static void tx_credit_callback(unsigned long data)
870{
871 struct xenvif *vif = (struct xenvif *)data;
872 tx_add_credit(vif);
873 xen_netbk_check_rx_xenvif(vif);
874}
875
876static void netbk_tx_err(struct xenvif *vif,
877 struct xen_netif_tx_request *txp, RING_IDX end)
878{
879 RING_IDX cons = vif->tx.req_cons;
880
881 do {
882 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
b9149729 883 if (cons == end)
f942dc25
IC		 884 break;
885 txp = RING_GET_REQUEST(&vif->tx, cons++);
886 } while (1);
887 vif->tx.req_cons = cons;
888 xen_netbk_check_rx_xenvif(vif);
889 xenvif_put(vif);
890}
891
48856286
IC		 892static void netbk_fatal_tx_err(struct xenvif *vif)
893{
894 netdev_err(vif->dev, "fatal error; disabling device\n");
895 xenvif_carrier_off(vif);
629821d9 896 xenvif_put(vif);
48856286
IC		 897}
898
f942dc25
IC		 899static int netbk_count_requests(struct xenvif *vif,
900 struct xen_netif_tx_request *first,
901 struct xen_netif_tx_request *txp,
902 int work_to_do)
903{
904 RING_IDX cons = vif->tx.req_cons;
905 int frags = 0;
906
907 if (!(first->flags & XEN_NETTXF_more_data))
908 return 0;
909
910 do {
911 if (frags >= work_to_do) {
48856286
IC		 912 netdev_err(vif->dev, "Need more frags\n");
913 netbk_fatal_tx_err(vif);
35876b5f 914 return -ENODATA;
f942dc25
IC		 915 }
916
917 if (unlikely(frags >= MAX_SKB_FRAGS)) {
48856286
IC		 918 netdev_err(vif->dev, "Too many frags\n");
919 netbk_fatal_tx_err(vif);
35876b5f 920 return -E2BIG;
f942dc25
IC		 921 }
922
923 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + frags),
924 sizeof(*txp));
925 if (txp->size > first->size) {
48856286
IC		 926 netdev_err(vif->dev, "Frag is bigger than frame.\n");
927 netbk_fatal_tx_err(vif);
35876b5f 928 return -EIO;
f942dc25
IC		 929 }
930
931 first->size -= txp->size;
932 frags++;
933
934 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
48856286 935 netdev_err(vif->dev, "txp->offset: %x, size: %u\n",
f942dc25 936 txp->offset, txp->size);
48856286 937 netbk_fatal_tx_err(vif);
35876b5f 938 return -EINVAL;
f942dc25
IC		 939 }
940 } while ((txp++)->flags & XEN_NETTXF_more_data);
941 return frags;
942}
943
944static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
ea066ad1 945 u16 pending_idx)
f942dc25
IC		 946{
947 struct page *page;
948 page = alloc_page(GFP_KERNEL|__GFP_COLD);
949 if (!page)
950 return NULL;
951 set_page_ext(page, netbk, pending_idx);
952 netbk->mmap_pages[pending_idx] = page;
953 return page;
954}
955
956static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
957 struct xenvif *vif,
958 struct sk_buff *skb,
959 struct xen_netif_tx_request *txp,
960 struct gnttab_copy *gop)
961{
962 struct skb_shared_info *shinfo = skb_shinfo(skb);
963 skb_frag_t *frags = shinfo->frags;
ea066ad1 964 u16 pending_idx = *((u16 *)skb->data);
f942dc25
IC		 965 int i, start;
966
967 /* Skip first skb fragment if it is on same page as header fragment. */
ea066ad1 968 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
f942dc25
IC		 969
970 for (i = start; i < shinfo->nr_frags; i++, txp++) {
971 struct page *page;
972 pending_ring_idx_t index;
973 struct pending_tx_info *pending_tx_info =
974 netbk->pending_tx_info;
975
976 index = pending_index(netbk->pending_cons++);
977 pending_idx = netbk->pending_ring[index];
27f85228 978 page = xen_netbk_alloc_page(netbk, pending_idx);
f942dc25 979 if (!page)
4cc7c1cb 980 goto err;
f942dc25 981
f942dc25
IC		 982 gop->source.u.ref = txp->gref;
983 gop->source.domid = vif->domid;
984 gop->source.offset = txp->offset;
985
986 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
987 gop->dest.domid = DOMID_SELF;
988 gop->dest.offset = txp->offset;
989
990 gop->len = txp->size;
991 gop->flags = GNTCOPY_source_gref;
992
993 gop++;
994
995 memcpy(&pending_tx_info[pending_idx].req, txp, sizeof(*txp));
996 xenvif_get(vif);
997 pending_tx_info[pending_idx].vif = vif;
ea066ad1 998 frag_set_pending_idx(&frags[i], pending_idx);
f942dc25
IC		 999 }
1000
1001 return gop;
4cc7c1cb
IC		 1002err:
1003 /* Unwind, freeing all pages and sending error responses. */
1004 while (i-- > start) {
1005 xen_netbk_idx_release(netbk, frag_get_pending_idx(&frags[i]),
1006 XEN_NETIF_RSP_ERROR);
1007 }
1008 /* The head too, if necessary. */
1009 if (start)
1010 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
1011
1012 return NULL;
f942dc25
IC		 1013}
1014
1015static int xen_netbk_tx_check_gop(struct xen_netbk *netbk,
1016 struct sk_buff *skb,
1017 struct gnttab_copy **gopp)
1018{
1019 struct gnttab_copy *gop = *gopp;
ea066ad1 1020 u16 pending_idx = *((u16 *)skb->data);
f942dc25
IC		 1021 struct skb_shared_info *shinfo = skb_shinfo(skb);
1022 int nr_frags = shinfo->nr_frags;
1023 int i, err, start;
1024
1025 /* Check status of header. */
1026 err = gop->status;
7d5145d8
MD		 1027 if (unlikely(err))
1028 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
f942dc25
IC		 1029
1030 /* Skip first skb fragment if it is on same page as header fragment. */
ea066ad1 1031 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
f942dc25
IC		 1032
1033 for (i = start; i < nr_frags; i++) {
1034 int j, newerr;
f942dc25 1035
ea066ad1 1036 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
f942dc25
IC		 1037
1038 /* Check error status: if okay then remember grant handle. */
1039 newerr = (++gop)->status;
1040 if (likely(!newerr)) {
1041 /* Had a previous error? Invalidate this fragment. */
1042 if (unlikely(err))
7d5145d8 1043 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
f942dc25
IC		 1044 continue;
1045 }
1046
1047 /* Error on this fragment: respond to client with an error. */
7d5145d8 1048 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_ERROR);
f942dc25
IC		 1049
1050 /* Not the first error? Preceding frags already invalidated. */
1051 if (err)
1052 continue;
1053
1054 /* First error: invalidate header and preceding fragments. */
1055 pending_idx = *((u16 *)skb->data);
7d5145d8 1056 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
f942dc25 1057 for (j = start; j < i; j++) {
5ccb3ea7 1058 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
7d5145d8 1059 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
f942dc25
IC		 1060 }
1061
1062 /* Remember the error: invalidate all subsequent fragments. */
1063 err = newerr;
1064 }
1065
1066 *gopp = gop + 1;
1067 return err;
1068}
1069
1070static void xen_netbk_fill_frags(struct xen_netbk *netbk, struct sk_buff *skb)
1071{
1072 struct skb_shared_info *shinfo = skb_shinfo(skb);
1073 int nr_frags = shinfo->nr_frags;
1074 int i;
1075
1076 for (i = 0; i < nr_frags; i++) {
1077 skb_frag_t *frag = shinfo->frags + i;
1078 struct xen_netif_tx_request *txp;
ea066ad1
IC		 1079 struct page *page;
1080 u16 pending_idx;
f942dc25 1081
ea066ad1 1082 pending_idx = frag_get_pending_idx(frag);
f942dc25
IC		 1083
1084 txp = &netbk->pending_tx_info[pending_idx].req;
ea066ad1
IC		 1085 page = virt_to_page(idx_to_kaddr(netbk, pending_idx));
1086 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
f942dc25
IC		 1087 skb->len += txp->size;
1088 skb->data_len += txp->size;
1089 skb->truesize += txp->size;
1090
1091 /* Take an extra reference to offset xen_netbk_idx_release */
1092 get_page(netbk->mmap_pages[pending_idx]);
7d5145d8 1093 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
f942dc25
IC		 1094 }
1095}
1096
1097static int xen_netbk_get_extras(struct xenvif *vif,
1098 struct xen_netif_extra_info *extras,
1099 int work_to_do)
1100{
1101 struct xen_netif_extra_info extra;
1102 RING_IDX cons = vif->tx.req_cons;
1103
1104 do {
1105 if (unlikely(work_to_do-- <= 0)) {
48856286
IC		 1106 netdev_err(vif->dev, "Missing extra info\n");
1107 netbk_fatal_tx_err(vif);
f942dc25
IC		 1108 return -EBADR;
1109 }
1110
1111 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1112 sizeof(extra));
1113 if (unlikely(!extra.type ||
1114 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1115 vif->tx.req_cons = ++cons;
48856286 1116 netdev_err(vif->dev,
f942dc25 1117 "Invalid extra type: %d\n", extra.type);
48856286 1118 netbk_fatal_tx_err(vif);
f942dc25
IC		 1119 return -EINVAL;
1120 }
1121
1122 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1123 vif->tx.req_cons = ++cons;
1124 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1125
1126 return work_to_do;
1127}
1128
1129static int netbk_set_skb_gso(struct xenvif *vif,
1130 struct sk_buff *skb,
1131 struct xen_netif_extra_info *gso)
1132{
1133 if (!gso->u.gso.size) {
48856286
IC		 1134 netdev_err(vif->dev, "GSO size must not be zero.\n");
1135 netbk_fatal_tx_err(vif);
f942dc25
IC		 1136 return -EINVAL;
1137 }
1138
1139 /* Currently only TCPv4 S.O. is supported. */
1140 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
48856286
IC		 1141 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1142 netbk_fatal_tx_err(vif);
f942dc25
IC		 1143 return -EINVAL;
1144 }
1145
1146 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1147 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1148
1149 /* Header must be checked, and gso_segs computed. */
1150 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1151 skb_shinfo(skb)->gso_segs = 0;
1152
1153 return 0;
1154}
1155
1156static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1157{
1158 struct iphdr *iph;
f942dc25
IC		 1159 int err = -EPROTO;
1160 int recalculate_partial_csum = 0;
1161
1162 /*
1163 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1164 * peers can fail to set NETRXF_csum_blank when sending a GSO
1165 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1166 * recalculate the partial checksum.
1167 */
1168 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1169 vif->rx_gso_checksum_fixup++;
1170 skb->ip_summed = CHECKSUM_PARTIAL;
1171 recalculate_partial_csum = 1;
1172 }
1173
1174 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1175 if (skb->ip_summed != CHECKSUM_PARTIAL)
1176 return 0;
1177
1178 if (skb->protocol != htons(ETH_P_IP))
1179 goto out;
1180
1181 iph = (void *)skb->data;
f942dc25
IC		 1182 switch (iph->protocol) {
1183 case IPPROTO_TCP:
bea89336
JW		 1184 if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1185 offsetof(struct tcphdr, check)))
1186 goto out;
f942dc25
IC		 1187
1188 if (recalculate_partial_csum) {
bea89336 1189 struct tcphdr *tcph = tcp_hdr(skb);
f942dc25
IC		 1190 tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1191 skb->len - iph->ihl*4,
1192 IPPROTO_TCP, 0);
1193 }
1194 break;
1195 case IPPROTO_UDP:
bea89336
JW		 1196 if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1197 offsetof(struct udphdr, check)))
1198 goto out;
f942dc25
IC		 1199
1200 if (recalculate_partial_csum) {
bea89336 1201 struct udphdr *udph = udp_hdr(skb);
f942dc25
IC		 1202 udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1203 skb->len - iph->ihl*4,
1204 IPPROTO_UDP, 0);
1205 }
1206 break;
1207 default:
1208 if (net_ratelimit())
1209 netdev_err(vif->dev,
1210 "Attempting to checksum a non-TCP/UDP packet, dropping a protocol %d packet\n",
1211 iph->protocol);
1212 goto out;
1213 }
1214
f942dc25
IC		 1215 err = 0;
1216
1217out:
1218 return err;
1219}
1220
1221static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1222{
1223 unsigned long now = jiffies;
1224 unsigned long next_credit =
1225 vif->credit_timeout.expires +
1226 msecs_to_jiffies(vif->credit_usec / 1000);
1227
1228 /* Timer could already be pending in rare cases. */
1229 if (timer_pending(&vif->credit_timeout))
1230 return true;
1231
1232 /* Passed the point where we can replenish credit? */
1233 if (time_after_eq(now, next_credit)) {
1234 vif->credit_timeout.expires = now;
1235 tx_add_credit(vif);
1236 }
1237
1238 /* Still too big to send right now? Set a callback. */
1239 if (size > vif->remaining_credit) {
1240 vif->credit_timeout.data =
1241 (unsigned long)vif;
1242 vif->credit_timeout.function =
1243 tx_credit_callback;
1244 mod_timer(&vif->credit_timeout,
1245 next_credit);
1246
1247 return true;
1248 }
1249
1250 return false;
1251}
1252
1253static unsigned xen_netbk_tx_build_gops(struct xen_netbk *netbk)
1254{
1255 struct gnttab_copy *gop = netbk->tx_copy_ops, *request_gop;
1256 struct sk_buff *skb;
1257 int ret;
1258
1259 while (((nr_pending_reqs(netbk) + MAX_SKB_FRAGS) < MAX_PENDING_REQS) &&
1260 !list_empty(&netbk->net_schedule_list)) {
1261 struct xenvif *vif;
1262 struct xen_netif_tx_request txreq;
1263 struct xen_netif_tx_request txfrags[MAX_SKB_FRAGS];
1264 struct page *page;
1265 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1266 u16 pending_idx;
1267 RING_IDX idx;
1268 int work_to_do;
1269 unsigned int data_len;
1270 pending_ring_idx_t index;
1271
1272 /* Get a netif from the list with work to do. */
1273 vif = poll_net_schedule_list(netbk);
48856286
IC		 1274 /* This can sometimes happen because the test of
1275 * list_empty(net_schedule_list) at the top of the
1276 * loop is unlocked. Just go back and have another
1277 * look.
1278 */
f942dc25
IC		 1279 if (!vif)
1280 continue;
1281
48856286
IC		 1282 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1283 XEN_NETIF_TX_RING_SIZE) {
1284 netdev_err(vif->dev,
1285 "Impossible number of requests. "
1286 "req_prod %d, req_cons %d, size %ld\n",
1287 vif->tx.sring->req_prod, vif->tx.req_cons,
1288 XEN_NETIF_TX_RING_SIZE);
1289 netbk_fatal_tx_err(vif);
1290 continue;
1291 }
1292
f942dc25
IC		 1293 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, work_to_do);
1294 if (!work_to_do) {
1295 xenvif_put(vif);
1296 continue;
1297 }
1298
1299 idx = vif->tx.req_cons;
1300 rmb(); /* Ensure that we see the request before we copy it. */
1301 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1302
1303 /* Credit-based scheduling. */
1304 if (txreq.size > vif->remaining_credit &&
1305 tx_credit_exceeded(vif, txreq.size)) {
1306 xenvif_put(vif);
1307 continue;
1308 }
1309
1310 vif->remaining_credit -= txreq.size;
1311
1312 work_to_do--;
1313 vif->tx.req_cons = ++idx;
1314
1315 memset(extras, 0, sizeof(extras));
1316 if (txreq.flags & XEN_NETTXF_extra_info) {
1317 work_to_do = xen_netbk_get_extras(vif, extras,
1318 work_to_do);
1319 idx = vif->tx.req_cons;
48856286 1320 if (unlikely(work_to_do < 0))
f942dc25 1321 continue;
f942dc25
IC		 1322 }
1323
1324 ret = netbk_count_requests(vif, &txreq, txfrags, work_to_do);
48856286 1325 if (unlikely(ret < 0))
f942dc25 1326 continue;
48856286 1327
f942dc25
IC		 1328 idx += ret;
1329
1330 if (unlikely(txreq.size < ETH_HLEN)) {
1331 netdev_dbg(vif->dev,
1332 "Bad packet size: %d\n", txreq.size);
1333 netbk_tx_err(vif, &txreq, idx);
1334 continue;
1335 }
1336
1337 /* No crossing a page as the payload mustn't fragment. */
1338 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
48856286 1339 netdev_err(vif->dev,
f942dc25
IC		 1340 "txreq.offset: %x, size: %u, end: %lu\n",
1341 txreq.offset, txreq.size,
1342 (txreq.offset&~PAGE_MASK) + txreq.size);
48856286 1343 netbk_fatal_tx_err(vif);
f942dc25
IC		 1344 continue;
1345 }
1346
1347 index = pending_index(netbk->pending_cons);
1348 pending_idx = netbk->pending_ring[index];
1349
1350 data_len = (txreq.size > PKT_PROT_LEN &&
1351 ret < MAX_SKB_FRAGS) ?
1352 PKT_PROT_LEN : txreq.size;
1353
1354 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1355 GFP_ATOMIC | __GFP_NOWARN);
1356 if (unlikely(skb == NULL)) {
1357 netdev_dbg(vif->dev,
1358 "Can't allocate a skb in start_xmit.\n");
1359 netbk_tx_err(vif, &txreq, idx);
1360 break;
1361 }
1362
1363 /* Packets passed to netif_rx() must have some headroom. */
1364 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1365
1366 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1367 struct xen_netif_extra_info *gso;
1368 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1369
1370 if (netbk_set_skb_gso(vif, skb, gso)) {
48856286 1371 /* Failure in netbk_set_skb_gso is fatal. */
f942dc25 1372 kfree_skb(skb);
f942dc25
IC		 1373 continue;
1374 }
1375 }
1376
1377 /* XXX could copy straight to head */
27f85228 1378 page = xen_netbk_alloc_page(netbk, pending_idx);
f942dc25
IC		 1379 if (!page) {
1380 kfree_skb(skb);
1381 netbk_tx_err(vif, &txreq, idx);
1382 continue;
1383 }
1384
f942dc25
IC		 1385 gop->source.u.ref = txreq.gref;
1386 gop->source.domid = vif->domid;
1387 gop->source.offset = txreq.offset;
1388
1389 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1390 gop->dest.domid = DOMID_SELF;
1391 gop->dest.offset = txreq.offset;
1392
1393 gop->len = txreq.size;
1394 gop->flags = GNTCOPY_source_gref;
1395
1396 gop++;
1397
1398 memcpy(&netbk->pending_tx_info[pending_idx].req,
1399 &txreq, sizeof(txreq));
1400 netbk->pending_tx_info[pending_idx].vif = vif;
1401 *((u16 *)skb->data) = pending_idx;
1402
1403 __skb_put(skb, data_len);
1404
1405 skb_shinfo(skb)->nr_frags = ret;
1406 if (data_len < txreq.size) {
1407 skb_shinfo(skb)->nr_frags++;
ea066ad1
IC		 1408 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1409 pending_idx);
f942dc25 1410 } else {
ea066ad1
IC		 1411 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1412 INVALID_PENDING_IDX);
f942dc25
IC		 1413 }
1414
f942dc25
IC		 1415 netbk->pending_cons++;
1416
1417 request_gop = xen_netbk_get_requests(netbk, vif,
1418 skb, txfrags, gop);
1419 if (request_gop == NULL) {
1420 kfree_skb(skb);
1421 netbk_tx_err(vif, &txreq, idx);
1422 continue;
1423 }
1424 gop = request_gop;
1425
1e0b6eac
AL		 1426 __skb_queue_tail(&netbk->tx_queue, skb);
1427
f942dc25
IC		 1428 vif->tx.req_cons = idx;
1429 xen_netbk_check_rx_xenvif(vif);
1430
1431 if ((gop-netbk->tx_copy_ops) >= ARRAY_SIZE(netbk->tx_copy_ops))
1432 break;
1433 }
1434
1435 return gop - netbk->tx_copy_ops;
1436}
1437
1438static void xen_netbk_tx_submit(struct xen_netbk *netbk)
1439{
1440 struct gnttab_copy *gop = netbk->tx_copy_ops;
1441 struct sk_buff *skb;
1442
1443 while ((skb = __skb_dequeue(&netbk->tx_queue)) != NULL) {
1444 struct xen_netif_tx_request *txp;
1445 struct xenvif *vif;
1446 u16 pending_idx;
1447 unsigned data_len;
1448
1449 pending_idx = *((u16 *)skb->data);
1450 vif = netbk->pending_tx_info[pending_idx].vif;
1451 txp = &netbk->pending_tx_info[pending_idx].req;
1452
1453 /* Check the remap error code. */
1454 if (unlikely(xen_netbk_tx_check_gop(netbk, skb, &gop))) {
1455 netdev_dbg(vif->dev, "netback grant failed.\n");
1456 skb_shinfo(skb)->nr_frags = 0;
1457 kfree_skb(skb);
1458 continue;
1459 }
1460
1461 data_len = skb->len;
1462 memcpy(skb->data,
1463 (void *)(idx_to_kaddr(netbk, pending_idx)|txp->offset),
1464 data_len);
1465 if (data_len < txp->size) {
1466 /* Append the packet payload as a fragment. */
1467 txp->offset += data_len;
1468 txp->size -= data_len;
1469 } else {
1470 /* Schedule a response immediately. */
7d5145d8 1471 xen_netbk_idx_release(netbk, pending_idx, XEN_NETIF_RSP_OKAY);
f942dc25
IC		 1472 }
1473
1474 if (txp->flags & XEN_NETTXF_csum_blank)
1475 skb->ip_summed = CHECKSUM_PARTIAL;
1476 else if (txp->flags & XEN_NETTXF_data_validated)
1477 skb->ip_summed = CHECKSUM_UNNECESSARY;
1478
1479 xen_netbk_fill_frags(netbk, skb);
1480
1481 /*
1482 * If the initial fragment was < PKT_PROT_LEN then
1483 * pull through some bytes from the other fragments to
1484 * increase the linear region to PKT_PROT_LEN bytes.
1485 */
1486 if (skb_headlen(skb) < PKT_PROT_LEN && skb_is_nonlinear(skb)) {
1487 int target = min_t(int, skb->len, PKT_PROT_LEN);
1488 __pskb_pull_tail(skb, target - skb_headlen(skb));
1489 }
1490
1491 skb->dev = vif->dev;
1492 skb->protocol = eth_type_trans(skb, skb->dev);
f9ca8f74 1493 skb_reset_network_header(skb);
f942dc25
IC		 1494
1495 if (checksum_setup(vif, skb)) {
1496 netdev_dbg(vif->dev,
1497 "Can't setup checksum in net_tx_action\n");
1498 kfree_skb(skb);
1499 continue;
1500 }
1501
40893fd0 1502 skb_probe_transport_header(skb, 0);
f9ca8f74 1503
f942dc25
IC		 1504 vif->dev->stats.rx_bytes += skb->len;
1505 vif->dev->stats.rx_packets++;
1506
1507 xenvif_receive_skb(vif, skb);
1508 }
1509}
1510
1511/* Called after netfront has transmitted */
1512static void xen_netbk_tx_action(struct xen_netbk *netbk)
1513{
1514 unsigned nr_gops;
f942dc25
IC		 1515
1516 nr_gops = xen_netbk_tx_build_gops(netbk);
1517
1518 if (nr_gops == 0)
1519 return;
f942dc25 1520
c571898f 1521 gnttab_batch_copy(netbk->tx_copy_ops, nr_gops);
f942dc25 1522
c571898f 1523 xen_netbk_tx_submit(netbk);
f942dc25
IC		 1524}
1525
7d5145d8
MD		 1526static void xen_netbk_idx_release(struct xen_netbk *netbk, u16 pending_idx,
1527 u8 status)
f942dc25
IC		 1528{
1529 struct xenvif *vif;
1530 struct pending_tx_info *pending_tx_info;
1531 pending_ring_idx_t index;
1532
1533 /* Already complete? */
1534 if (netbk->mmap_pages[pending_idx] == NULL)
1535 return;
1536
1537 pending_tx_info = &netbk->pending_tx_info[pending_idx];
1538
1539 vif = pending_tx_info->vif;
1540
7d5145d8 1541 make_tx_response(vif, &pending_tx_info->req, status);
f942dc25
IC		 1542
1543 index = pending_index(netbk->pending_prod++);
1544 netbk->pending_ring[index] = pending_idx;
1545
1546 xenvif_put(vif);
1547
9eaee8be 1548 netbk->mmap_pages[pending_idx]->mapping = NULL;
f942dc25
IC		 1549 put_page(netbk->mmap_pages[pending_idx]);
1550 netbk->mmap_pages[pending_idx] = NULL;
1551}
1552
1553static void make_tx_response(struct xenvif *vif,
1554 struct xen_netif_tx_request *txp,
1555 s8 st)
1556{
1557 RING_IDX i = vif->tx.rsp_prod_pvt;
1558 struct xen_netif_tx_response *resp;
1559 int notify;
1560
1561 resp = RING_GET_RESPONSE(&vif->tx, i);
1562 resp->id = txp->id;
1563 resp->status = st;
1564
1565 if (txp->flags & XEN_NETTXF_extra_info)
1566 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1567
1568 vif->tx.rsp_prod_pvt = ++i;
1569 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1570 if (notify)
1571 notify_remote_via_irq(vif->irq);
1572}
1573
1574static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1575 u16 id,
1576 s8 st,
1577 u16 offset,
1578 u16 size,
1579 u16 flags)
1580{
1581 RING_IDX i = vif->rx.rsp_prod_pvt;
1582 struct xen_netif_rx_response *resp;
1583
1584 resp = RING_GET_RESPONSE(&vif->rx, i);
1585 resp->offset = offset;
1586 resp->flags = flags;
1587 resp->id = id;
1588 resp->status = (s16)size;
1589 if (st < 0)
1590 resp->status = (s16)st;
1591
1592 vif->rx.rsp_prod_pvt = ++i;
1593
1594 return resp;
1595}
1596
1597static inline int rx_work_todo(struct xen_netbk *netbk)
1598{
1599 return !skb_queue_empty(&netbk->rx_queue);
1600}
1601
1602static inline int tx_work_todo(struct xen_netbk *netbk)
1603{
1604
1605 if (((nr_pending_reqs(netbk) + MAX_SKB_FRAGS) < MAX_PENDING_REQS) &&
1606 !list_empty(&netbk->net_schedule_list))
1607 return 1;
1608
1609 return 0;
1610}
1611
1612static int xen_netbk_kthread(void *data)
1613{
1614 struct xen_netbk *netbk = data;
1615 while (!kthread_should_stop()) {
1616 wait_event_interruptible(netbk->wq,
1617 rx_work_todo(netbk) ||
1618 tx_work_todo(netbk) ||
1619 kthread_should_stop());
1620 cond_resched();
1621
1622 if (kthread_should_stop())
1623 break;
1624
1625 if (rx_work_todo(netbk))
1626 xen_netbk_rx_action(netbk);
1627
1628 if (tx_work_todo(netbk))
1629 xen_netbk_tx_action(netbk);
1630 }
1631
1632 return 0;
1633}
1634
1635void xen_netbk_unmap_frontend_rings(struct xenvif *vif)
1636{
c9d63699
DV		 1637 if (vif->tx.sring)
1638 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1639 vif->tx.sring);
1640 if (vif->rx.sring)
1641 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1642 vif->rx.sring);
f942dc25
IC		 1643}
1644
1645int xen_netbk_map_frontend_rings(struct xenvif *vif,
1646 grant_ref_t tx_ring_ref,
1647 grant_ref_t rx_ring_ref)
1648{
c9d63699 1649 void *addr;
f942dc25
IC		 1650 struct xen_netif_tx_sring *txs;
1651 struct xen_netif_rx_sring *rxs;
1652
1653 int err = -ENOMEM;
1654
c9d63699
DV		 1655 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1656 tx_ring_ref, &addr);
1657 if (err)
f942dc25
IC		 1658 goto err;
1659
c9d63699 1660 txs = (struct xen_netif_tx_sring *)addr;
f942dc25
IC		 1661 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1662
c9d63699
DV		 1663 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1664 rx_ring_ref, &addr);
1665 if (err)
f942dc25 1666 goto err;
f942dc25 1667
c9d63699 1668 rxs = (struct xen_netif_rx_sring *)addr;
f942dc25
IC		 1669 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1670
c9d63699
DV		 1671 vif->rx_req_cons_peek = 0;
1672
f942dc25
IC		 1673 return 0;
1674
1675err:
1676 xen_netbk_unmap_frontend_rings(vif);
1677 return err;
1678}
1679
1680static int __init netback_init(void)
1681{
1682 int i;
1683 int rc = 0;
1684 int group;
1685
2a14b244 1686 if (!xen_domain())
f942dc25
IC		 1687 return -ENODEV;
1688
1689 xen_netbk_group_nr = num_online_cpus();
1690 xen_netbk = vzalloc(sizeof(struct xen_netbk) * xen_netbk_group_nr);
e404decb 1691 if (!xen_netbk)
f942dc25 1692 return -ENOMEM;
f942dc25
IC		 1693
1694 for (group = 0; group < xen_netbk_group_nr; group++) {
1695 struct xen_netbk *netbk = &xen_netbk[group];
1696 skb_queue_head_init(&netbk->rx_queue);
1697 skb_queue_head_init(&netbk->tx_queue);
1698
1699 init_timer(&netbk->net_timer);
1700 netbk->net_timer.data = (unsigned long)netbk;
1701 netbk->net_timer.function = xen_netbk_alarm;
1702
1703 netbk->pending_cons = 0;
1704 netbk->pending_prod = MAX_PENDING_REQS;
1705 for (i = 0; i < MAX_PENDING_REQS; i++)
1706 netbk->pending_ring[i] = i;
1707
1708 init_waitqueue_head(&netbk->wq);
1709 netbk->task = kthread_create(xen_netbk_kthread,
1710 (void *)netbk,
1711 "netback/%u", group);
1712
1713 if (IS_ERR(netbk->task)) {
6b84bd16 1714 printk(KERN_ALERT "kthread_create() fails at netback\n");
f942dc25
IC		 1715 del_timer(&netbk->net_timer);
1716 rc = PTR_ERR(netbk->task);
1717 goto failed_init;
1718 }
1719
1720 kthread_bind(netbk->task, group);
1721
1722 INIT_LIST_HEAD(&netbk->net_schedule_list);
1723
1724 spin_lock_init(&netbk->net_schedule_list_lock);
1725
1726 atomic_set(&netbk->netfront_count, 0);
1727
1728 wake_up_process(netbk->task);
1729 }
1730
1731 rc = xenvif_xenbus_init();
1732 if (rc)
1733 goto failed_init;
1734
1735 return 0;
1736
1737failed_init:
1738 while (--group >= 0) {
1739 struct xen_netbk *netbk = &xen_netbk[group];
1740 for (i = 0; i < MAX_PENDING_REQS; i++) {
1741 if (netbk->mmap_pages[i])
1742 __free_page(netbk->mmap_pages[i]);
1743 }
1744 del_timer(&netbk->net_timer);
1745 kthread_stop(netbk->task);
1746 }
1747 vfree(xen_netbk);
1748 return rc;
1749
1750}
1751
1752module_init(netback_init);
1753
1754MODULE_LICENSE("Dual BSD/GPL");
f984cec6 1755MODULE_ALIAS("xen-backend:vif");
Linux 6.x block layer and io_uring tree(s)