projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
xen-netback: Proper printf format for ptrdiff_t is 't'.
[deliverable/linux.git] / drivers / net / xen-netback / netback.c
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 #include <linux/highmem.h>
41
42 #include <net/tcp.h>
43
44 #include <xen/xen.h>
45 #include <xen/events.h>
46 #include <xen/interface/memory.h>
47
48 #include <asm/xen/hypercall.h>
49 #include <asm/xen/page.h>
50
51 /* Provide an option to disable split event channels at load time as
52 * event channels are limited resource. Split event channels are
53 * enabled by default.
54 */
55 bool separate_tx_rx_irq = 1;
56 module_param(separate_tx_rx_irq, bool, 0644);
57
58 /* When guest ring is filled up, qdisc queues the packets for us, but we have
59 * to timeout them, otherwise other guests' packets can get stucked there
60 */
61 unsigned int rx_drain_timeout_msecs = 10000;
62 module_param(rx_drain_timeout_msecs, uint, 0444);
63 unsigned int rx_drain_timeout_jiffies;
64
65 /*
66 * This is the maximum slots a skb can have. If a guest sends a skb
67 * which exceeds this limit it is considered malicious.
68 */
69 #define FATAL_SKB_SLOTS_DEFAULT 20
70 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
71 module_param(fatal_skb_slots, uint, 0444);
72
73 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
74 u8 status);
75
76 static void make_tx_response(struct xenvif *vif,
77 struct xen_netif_tx_request *txp,
78 s8 st);
79
80 static inline int tx_work_todo(struct xenvif *vif);
81 static inline int rx_work_todo(struct xenvif *vif);
82
83 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
84 u16 id,
85 s8 st,
86 u16 offset,
87 u16 size,
88 u16 flags);
89
90 static inline unsigned long idx_to_pfn(struct xenvif *vif,
91 u16 idx)
92 {
93 return page_to_pfn(vif->mmap_pages[idx]);
94 }
95
96 static inline unsigned long idx_to_kaddr(struct xenvif *vif,
97 u16 idx)
98 {
99 return (unsigned long)pfn_to_kaddr(idx_to_pfn(vif, idx));
100 }
101
102 /* Find the containing VIF's structure from a pointer in pending_tx_info array
103 */
104 static inline struct xenvif* ubuf_to_vif(struct ubuf_info *ubuf)
105 {
106 u16 pending_idx = ubuf->desc;
107 struct pending_tx_info *temp =
108 container_of(ubuf, struct pending_tx_info, callback_struct);
109 return container_of(temp - pending_idx,
110 struct xenvif,
111 pending_tx_info[0]);
112 }
113
114 /* This is a miniumum size for the linear area to avoid lots of
115 * calls to __pskb_pull_tail() as we set up checksum offsets. The
116 * value 128 was chosen as it covers all IPv4 and most likely
117 * IPv6 headers.
118 */
119 #define PKT_PROT_LEN 128
120
121 static u16 frag_get_pending_idx(skb_frag_t *frag)
122 {
123 return (u16)frag->page_offset;
124 }
125
126 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
127 {
128 frag->page_offset = pending_idx;
129 }
130
131 static inline pending_ring_idx_t pending_index(unsigned i)
132 {
133 return i & (MAX_PENDING_REQS-1);
134 }
135
136 bool xenvif_rx_ring_slots_available(struct xenvif *vif, int needed)
137 {
138 RING_IDX prod, cons;
139
140 do {
141 prod = vif->rx.sring->req_prod;
142 cons = vif->rx.req_cons;
143
144 if (prod - cons >= needed)
145 return true;
146
147 vif->rx.sring->req_event = prod + 1;
148
149 /* Make sure event is visible before we check prod
150 * again.
151 */
152 mb();
153 } while (vif->rx.sring->req_prod != prod);
154
155 return false;
156 }
157
158 /*
159 * Returns true if we should start a new receive buffer instead of
160 * adding 'size' bytes to a buffer which currently contains 'offset'
161 * bytes.
162 */
163 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
164 {
165 /* simple case: we have completely filled the current buffer. */
166 if (offset == MAX_BUFFER_OFFSET)
167 return true;
168
169 /*
170 * complex case: start a fresh buffer if the current frag
171 * would overflow the current buffer but only if:
172 * (i) this frag would fit completely in the next buffer
173 * and (ii) there is already some data in the current buffer
174 * and (iii) this is not the head buffer.
175 *
176 * Where:
177 * - (i) stops us splitting a frag into two copies
178 * unless the frag is too large for a single buffer.
179 * - (ii) stops us from leaving a buffer pointlessly empty.
180 * - (iii) stops us leaving the first buffer
181 * empty. Strictly speaking this is already covered
182 * by (ii) but is explicitly checked because
183 * netfront relies on the first buffer being
184 * non-empty and can crash otherwise.
185 *
186 * This means we will effectively linearise small
187 * frags but do not needlessly split large buffers
188 * into multiple copies tend to give large frags their
189 * own buffers as before.
190 */
191 if ((offset + size > MAX_BUFFER_OFFSET) &&
192 (size <= MAX_BUFFER_OFFSET) && offset && !head)
193 return true;
194
195 return false;
196 }
197
198 struct netrx_pending_operations {
199 unsigned copy_prod, copy_cons;
200 unsigned meta_prod, meta_cons;
201 struct gnttab_copy *copy;
202 struct xenvif_rx_meta *meta;
203 int copy_off;
204 grant_ref_t copy_gref;
205 };
206
207 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif *vif,
208 struct netrx_pending_operations *npo)
209 {
210 struct xenvif_rx_meta *meta;
211 struct xen_netif_rx_request *req;
212
213 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
214
215 meta = npo->meta + npo->meta_prod++;
216 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
217 meta->gso_size = 0;
218 meta->size = 0;
219 meta->id = req->id;
220
221 npo->copy_off = 0;
222 npo->copy_gref = req->gref;
223
224 return meta;
225 }
226
227 /*
228 * Set up the grant operations for this fragment. If it's a flipping
229 * interface, we also set up the unmap request from here.
230 */
231 static void xenvif_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
232 struct netrx_pending_operations *npo,
233 struct page *page, unsigned long size,
234 unsigned long offset, int *head,
235 struct xenvif *foreign_vif,
236 grant_ref_t foreign_gref)
237 {
238 struct gnttab_copy *copy_gop;
239 struct xenvif_rx_meta *meta;
240 unsigned long bytes;
241 int gso_type = XEN_NETIF_GSO_TYPE_NONE;
242
243 /* Data must not cross a page boundary. */
244 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
245
246 meta = npo->meta + npo->meta_prod - 1;
247
248 /* Skip unused frames from start of page */
249 page += offset >> PAGE_SHIFT;
250 offset &= ~PAGE_MASK;
251
252 while (size > 0) {
253 BUG_ON(offset >= PAGE_SIZE);
254 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
255
256 bytes = PAGE_SIZE - offset;
257
258 if (bytes > size)
259 bytes = size;
260
261 if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
262 /*
263 * Netfront requires there to be some data in the head
264 * buffer.
265 */
266 BUG_ON(*head);
267
268 meta = get_next_rx_buffer(vif, npo);
269 }
270
271 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
272 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
273
274 copy_gop = npo->copy + npo->copy_prod++;
275 copy_gop->flags = GNTCOPY_dest_gref;
276 copy_gop->len = bytes;
277
278 if (foreign_vif) {
279 copy_gop->source.domid = foreign_vif->domid;
280 copy_gop->source.u.ref = foreign_gref;
281 copy_gop->flags |= GNTCOPY_source_gref;
282 } else {
283 copy_gop->source.domid = DOMID_SELF;
284 copy_gop->source.u.gmfn =
285 virt_to_mfn(page_address(page));
286 }
287 copy_gop->source.offset = offset;
288
289 copy_gop->dest.domid = vif->domid;
290 copy_gop->dest.offset = npo->copy_off;
291 copy_gop->dest.u.ref = npo->copy_gref;
292
293 npo->copy_off += bytes;
294 meta->size += bytes;
295
296 offset += bytes;
297 size -= bytes;
298
299 /* Next frame */
300 if (offset == PAGE_SIZE && size) {
301 BUG_ON(!PageCompound(page));
302 page++;
303 offset = 0;
304 }
305
306 /* Leave a gap for the GSO descriptor. */
307 if (skb_is_gso(skb)) {
308 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
309 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
310 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
311 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
312 }
313
314 if (*head && ((1 << gso_type) & vif->gso_mask))
315 vif->rx.req_cons++;
316
317 *head = 0; /* There must be something in this buffer now. */
318
319 }
320 }
321
322 /*
323 * Prepare an SKB to be transmitted to the frontend.
324 *
325 * This function is responsible for allocating grant operations, meta
326 * structures, etc.
327 *
328 * It returns the number of meta structures consumed. The number of
329 * ring slots used is always equal to the number of meta slots used
330 * plus the number of GSO descriptors used. Currently, we use either
331 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
332 * frontend-side LRO).
333 */
334 static int xenvif_gop_skb(struct sk_buff *skb,
335 struct netrx_pending_operations *npo)
336 {
337 struct xenvif *vif = netdev_priv(skb->dev);
338 int nr_frags = skb_shinfo(skb)->nr_frags;
339 int i;
340 struct xen_netif_rx_request *req;
341 struct xenvif_rx_meta *meta;
342 unsigned char *data;
343 int head = 1;
344 int old_meta_prod;
345 int gso_type;
346 struct ubuf_info *ubuf = skb_shinfo(skb)->destructor_arg;
347 grant_ref_t foreign_grefs[MAX_SKB_FRAGS];
348 struct xenvif *foreign_vif = NULL;
349
350 old_meta_prod = npo->meta_prod;
351
352 gso_type = XEN_NETIF_GSO_TYPE_NONE;
353 if (skb_is_gso(skb)) {
354 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
355 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
356 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
357 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
358 }
359
360 /* Set up a GSO prefix descriptor, if necessary */
361 if ((1 << gso_type) & vif->gso_prefix_mask) {
362 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
363 meta = npo->meta + npo->meta_prod++;
364 meta->gso_type = gso_type;
365 meta->gso_size = skb_shinfo(skb)->gso_size;
366 meta->size = 0;
367 meta->id = req->id;
368 }
369
370 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
371 meta = npo->meta + npo->meta_prod++;
372
373 if ((1 << gso_type) & vif->gso_mask) {
374 meta->gso_type = gso_type;
375 meta->gso_size = skb_shinfo(skb)->gso_size;
376 } else {
377 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
378 meta->gso_size = 0;
379 }
380
381 meta->size = 0;
382 meta->id = req->id;
383 npo->copy_off = 0;
384 npo->copy_gref = req->gref;
385
386 if ((skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY) &&
387 (ubuf->callback == &xenvif_zerocopy_callback)) {
388 int i = 0;
389 foreign_vif = ubuf_to_vif(ubuf);
390
391 do {
392 u16 pending_idx = ubuf->desc;
393 foreign_grefs[i++] =
394 foreign_vif->pending_tx_info[pending_idx].req.gref;
395 ubuf = (struct ubuf_info *) ubuf->ctx;
396 } while (ubuf);
397 }
398
399 data = skb->data;
400 while (data < skb_tail_pointer(skb)) {
401 unsigned int offset = offset_in_page(data);
402 unsigned int len = PAGE_SIZE - offset;
403
404 if (data + len > skb_tail_pointer(skb))
405 len = skb_tail_pointer(skb) - data;
406
407 xenvif_gop_frag_copy(vif, skb, npo,
408 virt_to_page(data), len, offset, &head,
409 NULL,
410 0);
411 data += len;
412 }
413
414 for (i = 0; i < nr_frags; i++) {
415 xenvif_gop_frag_copy(vif, skb, npo,
416 skb_frag_page(&skb_shinfo(skb)->frags[i]),
417 skb_frag_size(&skb_shinfo(skb)->frags[i]),
418 skb_shinfo(skb)->frags[i].page_offset,
419 &head,
420 foreign_vif,
421 foreign_grefs[i]);
422 }
423
424 return npo->meta_prod - old_meta_prod;
425 }
426
427 /*
428 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
429 * used to set up the operations on the top of
430 * netrx_pending_operations, which have since been done. Check that
431 * they didn't give any errors and advance over them.
432 */
433 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
434 struct netrx_pending_operations *npo)
435 {
436 struct gnttab_copy *copy_op;
437 int status = XEN_NETIF_RSP_OKAY;
438 int i;
439
440 for (i = 0; i < nr_meta_slots; i++) {
441 copy_op = npo->copy + npo->copy_cons++;
442 if (copy_op->status != GNTST_okay) {
443 netdev_dbg(vif->dev,
444 "Bad status %d from copy to DOM%d.\n",
445 copy_op->status, vif->domid);
446 status = XEN_NETIF_RSP_ERROR;
447 }
448 }
449
450 return status;
451 }
452
453 static void xenvif_add_frag_responses(struct xenvif *vif, int status,
454 struct xenvif_rx_meta *meta,
455 int nr_meta_slots)
456 {
457 int i;
458 unsigned long offset;
459
460 /* No fragments used */
461 if (nr_meta_slots <= 1)
462 return;
463
464 nr_meta_slots--;
465
466 for (i = 0; i < nr_meta_slots; i++) {
467 int flags;
468 if (i == nr_meta_slots - 1)
469 flags = 0;
470 else
471 flags = XEN_NETRXF_more_data;
472
473 offset = 0;
474 make_rx_response(vif, meta[i].id, status, offset,
475 meta[i].size, flags);
476 }
477 }
478
479 struct xenvif_rx_cb {
480 int meta_slots_used;
481 };
482
483 #define XENVIF_RX_CB(skb) ((struct xenvif_rx_cb *)(skb)->cb)
484
485 void xenvif_kick_thread(struct xenvif *vif)
486 {
487 wake_up(&vif->wq);
488 }
489
490 static void xenvif_rx_action(struct xenvif *vif)
491 {
492 s8 status;
493 u16 flags;
494 struct xen_netif_rx_response *resp;
495 struct sk_buff_head rxq;
496 struct sk_buff *skb;
497 LIST_HEAD(notify);
498 int ret;
499 unsigned long offset;
500 bool need_to_notify = false;
501
502 struct netrx_pending_operations npo = {
503 .copy = vif->grant_copy_op,
504 .meta = vif->meta,
505 };
506
507 skb_queue_head_init(&rxq);
508
509 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL) {
510 RING_IDX max_slots_needed;
511 int i;
512
513 /* We need a cheap worse case estimate for the number of
514 * slots we'll use.
515 */
516
517 max_slots_needed = DIV_ROUND_UP(offset_in_page(skb->data) +
518 skb_headlen(skb),
519 PAGE_SIZE);
520 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
521 unsigned int size;
522 size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
523 max_slots_needed += DIV_ROUND_UP(size, PAGE_SIZE);
524 }
525 if (skb_is_gso(skb) &&
526 (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4 ||
527 skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6))
528 max_slots_needed++;
529
530 /* If the skb may not fit then bail out now */
531 if (!xenvif_rx_ring_slots_available(vif, max_slots_needed)) {
532 skb_queue_head(&vif->rx_queue, skb);
533 need_to_notify = true;
534 vif->rx_last_skb_slots = max_slots_needed;
535 break;
536 } else
537 vif->rx_last_skb_slots = 0;
538
539 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo);
540 BUG_ON(XENVIF_RX_CB(skb)->meta_slots_used > max_slots_needed);
541
542 __skb_queue_tail(&rxq, skb);
543 }
544
545 BUG_ON(npo.meta_prod > ARRAY_SIZE(vif->meta));
546
547 if (!npo.copy_prod)
548 goto done;
549
550 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
551 gnttab_batch_copy(vif->grant_copy_op, npo.copy_prod);
552
553 while ((skb = __skb_dequeue(&rxq)) != NULL) {
554
555 if ((1 << vif->meta[npo.meta_cons].gso_type) &
556 vif->gso_prefix_mask) {
557 resp = RING_GET_RESPONSE(&vif->rx,
558 vif->rx.rsp_prod_pvt++);
559
560 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
561
562 resp->offset = vif->meta[npo.meta_cons].gso_size;
563 resp->id = vif->meta[npo.meta_cons].id;
564 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
565
566 npo.meta_cons++;
567 XENVIF_RX_CB(skb)->meta_slots_used--;
568 }
569
570
571 vif->dev->stats.tx_bytes += skb->len;
572 vif->dev->stats.tx_packets++;
573
574 status = xenvif_check_gop(vif,
575 XENVIF_RX_CB(skb)->meta_slots_used,
576 &npo);
577
578 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
579 flags = 0;
580 else
581 flags = XEN_NETRXF_more_data;
582
583 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
584 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
585 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
586 /* remote but checksummed. */
587 flags |= XEN_NETRXF_data_validated;
588
589 offset = 0;
590 resp = make_rx_response(vif, vif->meta[npo.meta_cons].id,
591 status, offset,
592 vif->meta[npo.meta_cons].size,
593 flags);
594
595 if ((1 << vif->meta[npo.meta_cons].gso_type) &
596 vif->gso_mask) {
597 struct xen_netif_extra_info *gso =
598 (struct xen_netif_extra_info *)
599 RING_GET_RESPONSE(&vif->rx,
600 vif->rx.rsp_prod_pvt++);
601
602 resp->flags |= XEN_NETRXF_extra_info;
603
604 gso->u.gso.type = vif->meta[npo.meta_cons].gso_type;
605 gso->u.gso.size = vif->meta[npo.meta_cons].gso_size;
606 gso->u.gso.pad = 0;
607 gso->u.gso.features = 0;
608
609 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
610 gso->flags = 0;
611 }
612
613 xenvif_add_frag_responses(vif, status,
614 vif->meta + npo.meta_cons + 1,
615 XENVIF_RX_CB(skb)->meta_slots_used);
616
617 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
618
619 need_to_notify |= !!ret;
620
621 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
622 dev_kfree_skb(skb);
623 }
624
625 done:
626 if (need_to_notify)
627 notify_remote_via_irq(vif->rx_irq);
628 }
629
630 void xenvif_check_rx_xenvif(struct xenvif *vif)
631 {
632 int more_to_do;
633
634 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
635
636 if (more_to_do)
637 napi_schedule(&vif->napi);
638 }
639
640 static void tx_add_credit(struct xenvif *vif)
641 {
642 unsigned long max_burst, max_credit;
643
644 /*
645 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
646 * Otherwise the interface can seize up due to insufficient credit.
647 */
648 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
649 max_burst = min(max_burst, 131072UL);
650 max_burst = max(max_burst, vif->credit_bytes);
651
652 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
653 max_credit = vif->remaining_credit + vif->credit_bytes;
654 if (max_credit < vif->remaining_credit)
655 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
656
657 vif->remaining_credit = min(max_credit, max_burst);
658 }
659
660 static void tx_credit_callback(unsigned long data)
661 {
662 struct xenvif *vif = (struct xenvif *)data;
663 tx_add_credit(vif);
664 xenvif_check_rx_xenvif(vif);
665 }
666
667 static void xenvif_tx_err(struct xenvif *vif,
668 struct xen_netif_tx_request *txp, RING_IDX end)
669 {
670 RING_IDX cons = vif->tx.req_cons;
671 unsigned long flags;
672
673 do {
674 spin_lock_irqsave(&vif->response_lock, flags);
675 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
676 spin_unlock_irqrestore(&vif->response_lock, flags);
677 if (cons == end)
678 break;
679 txp = RING_GET_REQUEST(&vif->tx, cons++);
680 } while (1);
681 vif->tx.req_cons = cons;
682 }
683
684 static void xenvif_fatal_tx_err(struct xenvif *vif)
685 {
686 netdev_err(vif->dev, "fatal error; disabling device\n");
687 xenvif_carrier_off(vif);
688 }
689
690 static int xenvif_count_requests(struct xenvif *vif,
691 struct xen_netif_tx_request *first,
692 struct xen_netif_tx_request *txp,
693 int work_to_do)
694 {
695 RING_IDX cons = vif->tx.req_cons;
696 int slots = 0;
697 int drop_err = 0;
698 int more_data;
699
700 if (!(first->flags & XEN_NETTXF_more_data))
701 return 0;
702
703 do {
704 struct xen_netif_tx_request dropped_tx = { 0 };
705
706 if (slots >= work_to_do) {
707 netdev_err(vif->dev,
708 "Asked for %d slots but exceeds this limit\n",
709 work_to_do);
710 xenvif_fatal_tx_err(vif);
711 return -ENODATA;
712 }
713
714 /* This guest is really using too many slots and
715 * considered malicious.
716 */
717 if (unlikely(slots >= fatal_skb_slots)) {
718 netdev_err(vif->dev,
719 "Malicious frontend using %d slots, threshold %u\n",
720 slots, fatal_skb_slots);
721 xenvif_fatal_tx_err(vif);
722 return -E2BIG;
723 }
724
725 /* Xen network protocol had implicit dependency on
726 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
727 * the historical MAX_SKB_FRAGS value 18 to honor the
728 * same behavior as before. Any packet using more than
729 * 18 slots but less than fatal_skb_slots slots is
730 * dropped
731 */
732 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
733 if (net_ratelimit())
734 netdev_dbg(vif->dev,
735 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
736 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
737 drop_err = -E2BIG;
738 }
739
740 if (drop_err)
741 txp = &dropped_tx;
742
743 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
744 sizeof(*txp));
745
746 /* If the guest submitted a frame >= 64 KiB then
747 * first->size overflowed and following slots will
748 * appear to be larger than the frame.
749 *
750 * This cannot be fatal error as there are buggy
751 * frontends that do this.
752 *
753 * Consume all slots and drop the packet.
754 */
755 if (!drop_err && txp->size > first->size) {
756 if (net_ratelimit())
757 netdev_dbg(vif->dev,
758 "Invalid tx request, slot size %u > remaining size %u\n",
759 txp->size, first->size);
760 drop_err = -EIO;
761 }
762
763 first->size -= txp->size;
764 slots++;
765
766 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
767 netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
768 txp->offset, txp->size);
769 xenvif_fatal_tx_err(vif);
770 return -EINVAL;
771 }
772
773 more_data = txp->flags & XEN_NETTXF_more_data;
774
775 if (!drop_err)
776 txp++;
777
778 } while (more_data);
779
780 if (drop_err) {
781 xenvif_tx_err(vif, first, cons + slots);
782 return drop_err;
783 }
784
785 return slots;
786 }
787
788
789 struct xenvif_tx_cb {
790 u16 pending_idx;
791 };
792
793 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
794
795 static inline void xenvif_tx_create_gop(struct xenvif *vif,
796 u16 pending_idx,
797 struct xen_netif_tx_request *txp,
798 struct gnttab_map_grant_ref *gop)
799 {
800 vif->pages_to_map[gop-vif->tx_map_ops] = vif->mmap_pages[pending_idx];
801 gnttab_set_map_op(gop, idx_to_kaddr(vif, pending_idx),
802 GNTMAP_host_map | GNTMAP_readonly,
803 txp->gref, vif->domid);
804
805 memcpy(&vif->pending_tx_info[pending_idx].req, txp,
806 sizeof(*txp));
807 }
808
809 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
810 {
811 struct sk_buff *skb =
812 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
813 GFP_ATOMIC | __GFP_NOWARN);
814 if (unlikely(skb == NULL))
815 return NULL;
816
817 /* Packets passed to netif_rx() must have some headroom. */
818 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
819
820 /* Initialize it here to avoid later surprises */
821 skb_shinfo(skb)->destructor_arg = NULL;
822
823 return skb;
824 }
825
826 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif *vif,
827 struct sk_buff *skb,
828 struct xen_netif_tx_request *txp,
829 struct gnttab_map_grant_ref *gop)
830 {
831 struct skb_shared_info *shinfo = skb_shinfo(skb);
832 skb_frag_t *frags = shinfo->frags;
833 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
834 int start;
835 pending_ring_idx_t index;
836 unsigned int nr_slots, frag_overflow = 0;
837
838 /* At this point shinfo->nr_frags is in fact the number of
839 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
840 */
841 if (shinfo->nr_frags > MAX_SKB_FRAGS) {
842 frag_overflow = shinfo->nr_frags - MAX_SKB_FRAGS;
843 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
844 shinfo->nr_frags = MAX_SKB_FRAGS;
845 }
846 nr_slots = shinfo->nr_frags;
847
848 /* Skip first skb fragment if it is on same page as header fragment. */
849 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
850
851 for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
852 shinfo->nr_frags++, txp++, gop++) {
853 index = pending_index(vif->pending_cons++);
854 pending_idx = vif->pending_ring[index];
855 xenvif_tx_create_gop(vif, pending_idx, txp, gop);
856 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
857 }
858
859 if (frag_overflow) {
860 struct sk_buff *nskb = xenvif_alloc_skb(0);
861 if (unlikely(nskb == NULL)) {
862 if (net_ratelimit())
863 netdev_err(vif->dev,
864 "Can't allocate the frag_list skb.\n");
865 return NULL;
866 }
867
868 shinfo = skb_shinfo(nskb);
869 frags = shinfo->frags;
870
871 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
872 shinfo->nr_frags++, txp++, gop++) {
873 index = pending_index(vif->pending_cons++);
874 pending_idx = vif->pending_ring[index];
875 xenvif_tx_create_gop(vif, pending_idx, txp, gop);
876 frag_set_pending_idx(&frags[shinfo->nr_frags],
877 pending_idx);
878 }
879
880 skb_shinfo(skb)->frag_list = nskb;
881 }
882
883 return gop;
884 }
885
886 static inline void xenvif_grant_handle_set(struct xenvif *vif,
887 u16 pending_idx,
888 grant_handle_t handle)
889 {
890 if (unlikely(vif->grant_tx_handle[pending_idx] !=
891 NETBACK_INVALID_HANDLE)) {
892 netdev_err(vif->dev,
893 "Trying to overwrite active handle! pending_idx: %x\n",
894 pending_idx);
895 BUG();
896 }
897 vif->grant_tx_handle[pending_idx] = handle;
898 }
899
900 static inline void xenvif_grant_handle_reset(struct xenvif *vif,
901 u16 pending_idx)
902 {
903 if (unlikely(vif->grant_tx_handle[pending_idx] ==
904 NETBACK_INVALID_HANDLE)) {
905 netdev_err(vif->dev,
906 "Trying to unmap invalid handle! pending_idx: %x\n",
907 pending_idx);
908 BUG();
909 }
910 vif->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
911 }
912
913 static int xenvif_tx_check_gop(struct xenvif *vif,
914 struct sk_buff *skb,
915 struct gnttab_map_grant_ref **gopp)
916 {
917 struct gnttab_map_grant_ref *gop = *gopp;
918 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
919 struct skb_shared_info *shinfo = skb_shinfo(skb);
920 struct pending_tx_info *tx_info;
921 int nr_frags = shinfo->nr_frags;
922 int i, err, start;
923 struct sk_buff *first_skb = NULL;
924
925 /* Check status of header. */
926 err = gop->status;
927 if (unlikely(err))
928 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
929 else
930 xenvif_grant_handle_set(vif, pending_idx , gop->handle);
931
932 /* Skip first skb fragment if it is on same page as header fragment. */
933 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
934
935 check_frags:
936 for (i = start; i < nr_frags; i++) {
937 int j, newerr;
938
939 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
940 tx_info = &vif->pending_tx_info[pending_idx];
941
942 /* Check error status: if okay then remember grant handle. */
943 newerr = (++gop)->status;
944
945 if (likely(!newerr)) {
946 xenvif_grant_handle_set(vif, pending_idx , gop->handle);
947 /* Had a previous error? Invalidate this fragment. */
948 if (unlikely(err))
949 xenvif_idx_unmap(vif, pending_idx);
950 continue;
951 }
952
953 /* Error on this fragment: respond to client with an error. */
954 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
955
956 /* Not the first error? Preceding frags already invalidated. */
957 if (err)
958 continue;
959 /* First error: invalidate header and preceding fragments. */
960 if (!first_skb)
961 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
962 else
963 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
964 xenvif_idx_unmap(vif, pending_idx);
965 for (j = start; j < i; j++) {
966 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
967 xenvif_idx_unmap(vif, pending_idx);
968 }
969
970 /* Remember the error: invalidate all subsequent fragments. */
971 err = newerr;
972 }
973
974 if (skb_has_frag_list(skb)) {
975 first_skb = skb;
976 skb = shinfo->frag_list;
977 shinfo = skb_shinfo(skb);
978 nr_frags = shinfo->nr_frags;
979 start = 0;
980
981 goto check_frags;
982 }
983
984 /* There was a mapping error in the frag_list skb. We have to unmap
985 * the first skb's frags
986 */
987 if (first_skb && err) {
988 int j;
989 shinfo = skb_shinfo(first_skb);
990 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
991 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
992 for (j = start; j < shinfo->nr_frags; j++) {
993 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
994 xenvif_idx_unmap(vif, pending_idx);
995 }
996 }
997
998 *gopp = gop + 1;
999 return err;
1000 }
1001
1002 static void xenvif_fill_frags(struct xenvif *vif, struct sk_buff *skb)
1003 {
1004 struct skb_shared_info *shinfo = skb_shinfo(skb);
1005 int nr_frags = shinfo->nr_frags;
1006 int i;
1007 u16 prev_pending_idx = INVALID_PENDING_IDX;
1008
1009 if (skb_shinfo(skb)->destructor_arg)
1010 prev_pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1011
1012 for (i = 0; i < nr_frags; i++) {
1013 skb_frag_t *frag = shinfo->frags + i;
1014 struct xen_netif_tx_request *txp;
1015 struct page *page;
1016 u16 pending_idx;
1017
1018 pending_idx = frag_get_pending_idx(frag);
1019
1020 /* If this is not the first frag, chain it to the previous*/
1021 if (unlikely(prev_pending_idx == INVALID_PENDING_IDX))
1022 skb_shinfo(skb)->destructor_arg =
1023 &vif->pending_tx_info[pending_idx].callback_struct;
1024 else if (likely(pending_idx != prev_pending_idx))
1025 vif->pending_tx_info[prev_pending_idx].callback_struct.ctx =
1026 &(vif->pending_tx_info[pending_idx].callback_struct);
1027
1028 vif->pending_tx_info[pending_idx].callback_struct.ctx = NULL;
1029 prev_pending_idx = pending_idx;
1030
1031 txp = &vif->pending_tx_info[pending_idx].req;
1032 page = virt_to_page(idx_to_kaddr(vif, pending_idx));
1033 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1034 skb->len += txp->size;
1035 skb->data_len += txp->size;
1036 skb->truesize += txp->size;
1037
1038 /* Take an extra reference to offset network stack's put_page */
1039 get_page(vif->mmap_pages[pending_idx]);
1040 }
1041 /* FIXME: __skb_fill_page_desc set this to true because page->pfmemalloc
1042 * overlaps with "index", and "mapping" is not set. I think mapping
1043 * should be set. If delivered to local stack, it would drop this
1044 * skb in sk_filter unless the socket has the right to use it.
1045 */
1046 skb->pfmemalloc = false;
1047 }
1048
1049 static int xenvif_get_extras(struct xenvif *vif,
1050 struct xen_netif_extra_info *extras,
1051 int work_to_do)
1052 {
1053 struct xen_netif_extra_info extra;
1054 RING_IDX cons = vif->tx.req_cons;
1055
1056 do {
1057 if (unlikely(work_to_do-- <= 0)) {
1058 netdev_err(vif->dev, "Missing extra info\n");
1059 xenvif_fatal_tx_err(vif);
1060 return -EBADR;
1061 }
1062
1063 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1064 sizeof(extra));
1065 if (unlikely(!extra.type ||
1066 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1067 vif->tx.req_cons = ++cons;
1068 netdev_err(vif->dev,
1069 "Invalid extra type: %d\n", extra.type);
1070 xenvif_fatal_tx_err(vif);
1071 return -EINVAL;
1072 }
1073
1074 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1075 vif->tx.req_cons = ++cons;
1076 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1077
1078 return work_to_do;
1079 }
1080
1081 static int xenvif_set_skb_gso(struct xenvif *vif,
1082 struct sk_buff *skb,
1083 struct xen_netif_extra_info *gso)
1084 {
1085 if (!gso->u.gso.size) {
1086 netdev_err(vif->dev, "GSO size must not be zero.\n");
1087 xenvif_fatal_tx_err(vif);
1088 return -EINVAL;
1089 }
1090
1091 switch (gso->u.gso.type) {
1092 case XEN_NETIF_GSO_TYPE_TCPV4:
1093 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1094 break;
1095 case XEN_NETIF_GSO_TYPE_TCPV6:
1096 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1097 break;
1098 default:
1099 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1100 xenvif_fatal_tx_err(vif);
1101 return -EINVAL;
1102 }
1103
1104 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1105 /* gso_segs will be calculated later */
1106
1107 return 0;
1108 }
1109
1110 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1111 {
1112 bool recalculate_partial_csum = false;
1113
1114 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1115 * peers can fail to set NETRXF_csum_blank when sending a GSO
1116 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1117 * recalculate the partial checksum.
1118 */
1119 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1120 vif->rx_gso_checksum_fixup++;
1121 skb->ip_summed = CHECKSUM_PARTIAL;
1122 recalculate_partial_csum = true;
1123 }
1124
1125 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1126 if (skb->ip_summed != CHECKSUM_PARTIAL)
1127 return 0;
1128
1129 return skb_checksum_setup(skb, recalculate_partial_csum);
1130 }
1131
1132 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1133 {
1134 u64 now = get_jiffies_64();
1135 u64 next_credit = vif->credit_window_start +
1136 msecs_to_jiffies(vif->credit_usec / 1000);
1137
1138 /* Timer could already be pending in rare cases. */
1139 if (timer_pending(&vif->credit_timeout))
1140 return true;
1141
1142 /* Passed the point where we can replenish credit? */
1143 if (time_after_eq64(now, next_credit)) {
1144 vif->credit_window_start = now;
1145 tx_add_credit(vif);
1146 }
1147
1148 /* Still too big to send right now? Set a callback. */
1149 if (size > vif->remaining_credit) {
1150 vif->credit_timeout.data =
1151 (unsigned long)vif;
1152 vif->credit_timeout.function =
1153 tx_credit_callback;
1154 mod_timer(&vif->credit_timeout,
1155 next_credit);
1156 vif->credit_window_start = next_credit;
1157
1158 return true;
1159 }
1160
1161 return false;
1162 }
1163
1164 static unsigned xenvif_tx_build_gops(struct xenvif *vif, int budget)
1165 {
1166 struct gnttab_map_grant_ref *gop = vif->tx_map_ops, *request_gop;
1167 struct sk_buff *skb;
1168 int ret;
1169
1170 while (xenvif_tx_pending_slots_available(vif) &&
1171 (skb_queue_len(&vif->tx_queue) < budget)) {
1172 struct xen_netif_tx_request txreq;
1173 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1174 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1175 u16 pending_idx;
1176 RING_IDX idx;
1177 int work_to_do;
1178 unsigned int data_len;
1179 pending_ring_idx_t index;
1180
1181 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1182 XEN_NETIF_TX_RING_SIZE) {
1183 netdev_err(vif->dev,
1184 "Impossible number of requests. "
1185 "req_prod %d, req_cons %d, size %ld\n",
1186 vif->tx.sring->req_prod, vif->tx.req_cons,
1187 XEN_NETIF_TX_RING_SIZE);
1188 xenvif_fatal_tx_err(vif);
1189 continue;
1190 }
1191
1192 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);
1193 if (!work_to_do)
1194 break;
1195
1196 idx = vif->tx.req_cons;
1197 rmb(); /* Ensure that we see the request before we copy it. */
1198 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1199
1200 /* Credit-based scheduling. */
1201 if (txreq.size > vif->remaining_credit &&
1202 tx_credit_exceeded(vif, txreq.size))
1203 break;
1204
1205 vif->remaining_credit -= txreq.size;
1206
1207 work_to_do--;
1208 vif->tx.req_cons = ++idx;
1209
1210 memset(extras, 0, sizeof(extras));
1211 if (txreq.flags & XEN_NETTXF_extra_info) {
1212 work_to_do = xenvif_get_extras(vif, extras,
1213 work_to_do);
1214 idx = vif->tx.req_cons;
1215 if (unlikely(work_to_do < 0))
1216 break;
1217 }
1218
1219 ret = xenvif_count_requests(vif, &txreq, txfrags, work_to_do);
1220 if (unlikely(ret < 0))
1221 break;
1222
1223 idx += ret;
1224
1225 if (unlikely(txreq.size < ETH_HLEN)) {
1226 netdev_dbg(vif->dev,
1227 "Bad packet size: %d\n", txreq.size);
1228 xenvif_tx_err(vif, &txreq, idx);
1229 break;
1230 }
1231
1232 /* No crossing a page as the payload mustn't fragment. */
1233 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1234 netdev_err(vif->dev,
1235 "txreq.offset: %x, size: %u, end: %lu\n",
1236 txreq.offset, txreq.size,
1237 (txreq.offset&~PAGE_MASK) + txreq.size);
1238 xenvif_fatal_tx_err(vif);
1239 break;
1240 }
1241
1242 index = pending_index(vif->pending_cons);
1243 pending_idx = vif->pending_ring[index];
1244
1245 data_len = (txreq.size > PKT_PROT_LEN &&
1246 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1247 PKT_PROT_LEN : txreq.size;
1248
1249 skb = xenvif_alloc_skb(data_len);
1250 if (unlikely(skb == NULL)) {
1251 netdev_dbg(vif->dev,
1252 "Can't allocate a skb in start_xmit.\n");
1253 xenvif_tx_err(vif, &txreq, idx);
1254 break;
1255 }
1256
1257 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1258 struct xen_netif_extra_info *gso;
1259 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1260
1261 if (xenvif_set_skb_gso(vif, skb, gso)) {
1262 /* Failure in xenvif_set_skb_gso is fatal. */
1263 kfree_skb(skb);
1264 break;
1265 }
1266 }
1267
1268 xenvif_tx_create_gop(vif, pending_idx, &txreq, gop);
1269
1270 gop++;
1271
1272 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1273
1274 __skb_put(skb, data_len);
1275
1276 skb_shinfo(skb)->nr_frags = ret;
1277 if (data_len < txreq.size) {
1278 skb_shinfo(skb)->nr_frags++;
1279 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1280 pending_idx);
1281 } else {
1282 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1283 INVALID_PENDING_IDX);
1284 }
1285
1286 vif->pending_cons++;
1287
1288 request_gop = xenvif_get_requests(vif, skb, txfrags, gop);
1289 if (request_gop == NULL) {
1290 kfree_skb(skb);
1291 xenvif_tx_err(vif, &txreq, idx);
1292 break;
1293 }
1294 gop = request_gop;
1295
1296 __skb_queue_tail(&vif->tx_queue, skb);
1297
1298 vif->tx.req_cons = idx;
1299
1300 if ((gop-vif->tx_map_ops) >= ARRAY_SIZE(vif->tx_map_ops))
1301 break;
1302 }
1303
1304 return gop - vif->tx_map_ops;
1305 }
1306
1307 /* Consolidate skb with a frag_list into a brand new one with local pages on
1308 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1309 */
1310 static int xenvif_handle_frag_list(struct xenvif *vif, struct sk_buff *skb)
1311 {
1312 unsigned int offset = skb_headlen(skb);
1313 skb_frag_t frags[MAX_SKB_FRAGS];
1314 int i;
1315 struct ubuf_info *uarg;
1316 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1317
1318 vif->tx_zerocopy_sent += 2;
1319 vif->tx_frag_overflow++;
1320
1321 xenvif_fill_frags(vif, nskb);
1322 /* Subtract frags size, we will correct it later */
1323 skb->truesize -= skb->data_len;
1324 skb->len += nskb->len;
1325 skb->data_len += nskb->len;
1326
1327 /* create a brand new frags array and coalesce there */
1328 for (i = 0; offset < skb->len; i++) {
1329 struct page *page;
1330 unsigned int len;
1331
1332 BUG_ON(i >= MAX_SKB_FRAGS);
1333 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
1334 if (!page) {
1335 int j;
1336 skb->truesize += skb->data_len;
1337 for (j = 0; j < i; j++)
1338 put_page(frags[j].page.p);
1339 return -ENOMEM;
1340 }
1341
1342 if (offset + PAGE_SIZE < skb->len)
1343 len = PAGE_SIZE;
1344 else
1345 len = skb->len - offset;
1346 if (skb_copy_bits(skb, offset, page_address(page), len))
1347 BUG();
1348
1349 offset += len;
1350 frags[i].page.p = page;
1351 frags[i].page_offset = 0;
1352 skb_frag_size_set(&frags[i], len);
1353 }
1354 /* swap out with old one */
1355 memcpy(skb_shinfo(skb)->frags,
1356 frags,
1357 i * sizeof(skb_frag_t));
1358 skb_shinfo(skb)->nr_frags = i;
1359 skb->truesize += i * PAGE_SIZE;
1360
1361 /* remove traces of mapped pages and frag_list */
1362 skb_frag_list_init(skb);
1363 uarg = skb_shinfo(skb)->destructor_arg;
1364 uarg->callback(uarg, true);
1365 skb_shinfo(skb)->destructor_arg = NULL;
1366
1367 skb_shinfo(nskb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
1368 kfree_skb(nskb);
1369
1370 return 0;
1371 }
1372
1373 static int xenvif_tx_submit(struct xenvif *vif)
1374 {
1375 struct gnttab_map_grant_ref *gop = vif->tx_map_ops;
1376 struct sk_buff *skb;
1377 int work_done = 0;
1378
1379 while ((skb = __skb_dequeue(&vif->tx_queue)) != NULL) {
1380 struct xen_netif_tx_request *txp;
1381 u16 pending_idx;
1382 unsigned data_len;
1383
1384 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1385 txp = &vif->pending_tx_info[pending_idx].req;
1386
1387 /* Check the remap error code. */
1388 if (unlikely(xenvif_tx_check_gop(vif, skb, &gop))) {
1389 netdev_dbg(vif->dev, "netback grant failed.\n");
1390 skb_shinfo(skb)->nr_frags = 0;
1391 kfree_skb(skb);
1392 continue;
1393 }
1394
1395 data_len = skb->len;
1396 memcpy(skb->data,
1397 (void *)(idx_to_kaddr(vif, pending_idx)|txp->offset),
1398 data_len);
1399 vif->pending_tx_info[pending_idx].callback_struct.ctx = NULL;
1400 if (data_len < txp->size) {
1401 /* Append the packet payload as a fragment. */
1402 txp->offset += data_len;
1403 txp->size -= data_len;
1404 skb_shinfo(skb)->destructor_arg =
1405 &vif->pending_tx_info[pending_idx].callback_struct;
1406 } else {
1407 /* Schedule a response immediately. */
1408 xenvif_idx_unmap(vif, pending_idx);
1409 }
1410
1411 if (txp->flags & XEN_NETTXF_csum_blank)
1412 skb->ip_summed = CHECKSUM_PARTIAL;
1413 else if (txp->flags & XEN_NETTXF_data_validated)
1414 skb->ip_summed = CHECKSUM_UNNECESSARY;
1415
1416 xenvif_fill_frags(vif, skb);
1417
1418 if (unlikely(skb_has_frag_list(skb))) {
1419 if (xenvif_handle_frag_list(vif, skb)) {
1420 if (net_ratelimit())
1421 netdev_err(vif->dev,
1422 "Not enough memory to consolidate frag_list!\n");
1423 skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
1424 kfree_skb(skb);
1425 continue;
1426 }
1427 }
1428
1429 if (skb_is_nonlinear(skb) && skb_headlen(skb) < PKT_PROT_LEN) {
1430 int target = min_t(int, skb->len, PKT_PROT_LEN);
1431 __pskb_pull_tail(skb, target - skb_headlen(skb));
1432 }
1433
1434 skb->dev = vif->dev;
1435 skb->protocol = eth_type_trans(skb, skb->dev);
1436 skb_reset_network_header(skb);
1437
1438 if (checksum_setup(vif, skb)) {
1439 netdev_dbg(vif->dev,
1440 "Can't setup checksum in net_tx_action\n");
1441 /* We have to set this flag to trigger the callback */
1442 if (skb_shinfo(skb)->destructor_arg)
1443 skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
1444 kfree_skb(skb);
1445 continue;
1446 }
1447
1448 skb_probe_transport_header(skb, 0);
1449
1450 /* If the packet is GSO then we will have just set up the
1451 * transport header offset in checksum_setup so it's now
1452 * straightforward to calculate gso_segs.
1453 */
1454 if (skb_is_gso(skb)) {
1455 int mss = skb_shinfo(skb)->gso_size;
1456 int hdrlen = skb_transport_header(skb) -
1457 skb_mac_header(skb) +
1458 tcp_hdrlen(skb);
1459
1460 skb_shinfo(skb)->gso_segs =
1461 DIV_ROUND_UP(skb->len - hdrlen, mss);
1462 }
1463
1464 vif->dev->stats.rx_bytes += skb->len;
1465 vif->dev->stats.rx_packets++;
1466
1467 work_done++;
1468
1469 /* Set this flag right before netif_receive_skb, otherwise
1470 * someone might think this packet already left netback, and
1471 * do a skb_copy_ubufs while we are still in control of the
1472 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1473 */
1474 if (skb_shinfo(skb)->destructor_arg) {
1475 skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
1476 vif->tx_zerocopy_sent++;
1477 }
1478
1479 netif_receive_skb(skb);
1480 }
1481
1482 return work_done;
1483 }
1484
1485 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1486 {
1487 unsigned long flags;
1488 pending_ring_idx_t index;
1489 struct xenvif *vif = ubuf_to_vif(ubuf);
1490
1491 /* This is the only place where we grab this lock, to protect callbacks
1492 * from each other.
1493 */
1494 spin_lock_irqsave(&vif->callback_lock, flags);
1495 do {
1496 u16 pending_idx = ubuf->desc;
1497 ubuf = (struct ubuf_info *) ubuf->ctx;
1498 BUG_ON(vif->dealloc_prod - vif->dealloc_cons >=
1499 MAX_PENDING_REQS);
1500 index = pending_index(vif->dealloc_prod);
1501 vif->dealloc_ring[index] = pending_idx;
1502 /* Sync with xenvif_tx_dealloc_action:
1503 * insert idx then incr producer.
1504 */
1505 smp_wmb();
1506 vif->dealloc_prod++;
1507 } while (ubuf);
1508 wake_up(&vif->dealloc_wq);
1509 spin_unlock_irqrestore(&vif->callback_lock, flags);
1510
1511 if (RING_HAS_UNCONSUMED_REQUESTS(&vif->tx) &&
1512 xenvif_tx_pending_slots_available(vif)) {
1513 local_bh_disable();
1514 napi_schedule(&vif->napi);
1515 local_bh_enable();
1516 }
1517
1518 if (likely(zerocopy_success))
1519 vif->tx_zerocopy_success++;
1520 else
1521 vif->tx_zerocopy_fail++;
1522 }
1523
1524 static inline void xenvif_tx_dealloc_action(struct xenvif *vif)
1525 {
1526 struct gnttab_unmap_grant_ref *gop;
1527 pending_ring_idx_t dc, dp;
1528 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1529 unsigned int i = 0;
1530
1531 dc = vif->dealloc_cons;
1532 gop = vif->tx_unmap_ops;
1533
1534 /* Free up any grants we have finished using */
1535 do {
1536 dp = vif->dealloc_prod;
1537
1538 /* Ensure we see all indices enqueued by all
1539 * xenvif_zerocopy_callback().
1540 */
1541 smp_rmb();
1542
1543 while (dc != dp) {
1544 BUG_ON(gop - vif->tx_unmap_ops > MAX_PENDING_REQS);
1545 pending_idx =
1546 vif->dealloc_ring[pending_index(dc++)];
1547
1548 pending_idx_release[gop-vif->tx_unmap_ops] =
1549 pending_idx;
1550 vif->pages_to_unmap[gop-vif->tx_unmap_ops] =
1551 vif->mmap_pages[pending_idx];
1552 gnttab_set_unmap_op(gop,
1553 idx_to_kaddr(vif, pending_idx),
1554 GNTMAP_host_map,
1555 vif->grant_tx_handle[pending_idx]);
1556 /* Btw. already unmapped? */
1557 xenvif_grant_handle_reset(vif, pending_idx);
1558 ++gop;
1559 }
1560
1561 } while (dp != vif->dealloc_prod);
1562
1563 vif->dealloc_cons = dc;
1564
1565 if (gop - vif->tx_unmap_ops > 0) {
1566 int ret;
1567 ret = gnttab_unmap_refs(vif->tx_unmap_ops,
1568 NULL,
1569 vif->pages_to_unmap,
1570 gop - vif->tx_unmap_ops);
1571 if (ret) {
1572 netdev_err(vif->dev, "Unmap fail: nr_ops %tx ret %d\n",
1573 gop - vif->tx_unmap_ops, ret);
1574 for (i = 0; i < gop - vif->tx_unmap_ops; ++i) {
1575 if (gop[i].status != GNTST_okay)
1576 netdev_err(vif->dev,
1577 " host_addr: %llx handle: %x status: %d\n",
1578 gop[i].host_addr,
1579 gop[i].handle,
1580 gop[i].status);
1581 }
1582 BUG();
1583 }
1584 }
1585
1586 for (i = 0; i < gop - vif->tx_unmap_ops; ++i)
1587 xenvif_idx_release(vif, pending_idx_release[i],
1588 XEN_NETIF_RSP_OKAY);
1589 }
1590
1591
1592 /* Called after netfront has transmitted */
1593 int xenvif_tx_action(struct xenvif *vif, int budget)
1594 {
1595 unsigned nr_gops;
1596 int work_done, ret;
1597
1598 if (unlikely(!tx_work_todo(vif)))
1599 return 0;
1600
1601 nr_gops = xenvif_tx_build_gops(vif, budget);
1602
1603 if (nr_gops == 0)
1604 return 0;
1605
1606 ret = gnttab_map_refs(vif->tx_map_ops,
1607 NULL,
1608 vif->pages_to_map,
1609 nr_gops);
1610 BUG_ON(ret);
1611
1612 work_done = xenvif_tx_submit(vif);
1613
1614 return work_done;
1615 }
1616
1617 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
1618 u8 status)
1619 {
1620 struct pending_tx_info *pending_tx_info;
1621 pending_ring_idx_t index;
1622 unsigned long flags;
1623
1624 pending_tx_info = &vif->pending_tx_info[pending_idx];
1625 spin_lock_irqsave(&vif->response_lock, flags);
1626 make_tx_response(vif, &pending_tx_info->req, status);
1627 index = pending_index(vif->pending_prod);
1628 vif->pending_ring[index] = pending_idx;
1629 /* TX shouldn't use the index before we give it back here */
1630 mb();
1631 vif->pending_prod++;
1632 spin_unlock_irqrestore(&vif->response_lock, flags);
1633 }
1634
1635
1636 static void make_tx_response(struct xenvif *vif,
1637 struct xen_netif_tx_request *txp,
1638 s8 st)
1639 {
1640 RING_IDX i = vif->tx.rsp_prod_pvt;
1641 struct xen_netif_tx_response *resp;
1642 int notify;
1643
1644 resp = RING_GET_RESPONSE(&vif->tx, i);
1645 resp->id = txp->id;
1646 resp->status = st;
1647
1648 if (txp->flags & XEN_NETTXF_extra_info)
1649 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1650
1651 vif->tx.rsp_prod_pvt = ++i;
1652 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1653 if (notify)
1654 notify_remote_via_irq(vif->tx_irq);
1655 }
1656
1657 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1658 u16 id,
1659 s8 st,
1660 u16 offset,
1661 u16 size,
1662 u16 flags)
1663 {
1664 RING_IDX i = vif->rx.rsp_prod_pvt;
1665 struct xen_netif_rx_response *resp;
1666
1667 resp = RING_GET_RESPONSE(&vif->rx, i);
1668 resp->offset = offset;
1669 resp->flags = flags;
1670 resp->id = id;
1671 resp->status = (s16)size;
1672 if (st < 0)
1673 resp->status = (s16)st;
1674
1675 vif->rx.rsp_prod_pvt = ++i;
1676
1677 return resp;
1678 }
1679
1680 void xenvif_idx_unmap(struct xenvif *vif, u16 pending_idx)
1681 {
1682 int ret;
1683 struct gnttab_unmap_grant_ref tx_unmap_op;
1684
1685 gnttab_set_unmap_op(&tx_unmap_op,
1686 idx_to_kaddr(vif, pending_idx),
1687 GNTMAP_host_map,
1688 vif->grant_tx_handle[pending_idx]);
1689 /* Btw. already unmapped? */
1690 xenvif_grant_handle_reset(vif, pending_idx);
1691
1692 ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1693 &vif->mmap_pages[pending_idx], 1);
1694 BUG_ON(ret);
1695
1696 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
1697 }
1698
1699 static inline int rx_work_todo(struct xenvif *vif)
1700 {
1701 return (!skb_queue_empty(&vif->rx_queue) &&
1702 xenvif_rx_ring_slots_available(vif, vif->rx_last_skb_slots)) ||
1703 vif->rx_queue_purge;
1704 }
1705
1706 static inline int tx_work_todo(struct xenvif *vif)
1707 {
1708
1709 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->tx)) &&
1710 xenvif_tx_pending_slots_available(vif))
1711 return 1;
1712
1713 return 0;
1714 }
1715
1716 static inline bool tx_dealloc_work_todo(struct xenvif *vif)
1717 {
1718 return vif->dealloc_cons != vif->dealloc_prod;
1719 }
1720
1721 void xenvif_unmap_frontend_rings(struct xenvif *vif)
1722 {
1723 if (vif->tx.sring)
1724 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1725 vif->tx.sring);
1726 if (vif->rx.sring)
1727 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1728 vif->rx.sring);
1729 }
1730
1731 int xenvif_map_frontend_rings(struct xenvif *vif,
1732 grant_ref_t tx_ring_ref,
1733 grant_ref_t rx_ring_ref)
1734 {
1735 void *addr;
1736 struct xen_netif_tx_sring *txs;
1737 struct xen_netif_rx_sring *rxs;
1738
1739 int err = -ENOMEM;
1740
1741 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1742 tx_ring_ref, &addr);
1743 if (err)
1744 goto err;
1745
1746 txs = (struct xen_netif_tx_sring *)addr;
1747 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1748
1749 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1750 rx_ring_ref, &addr);
1751 if (err)
1752 goto err;
1753
1754 rxs = (struct xen_netif_rx_sring *)addr;
1755 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1756
1757 return 0;
1758
1759 err:
1760 xenvif_unmap_frontend_rings(vif);
1761 return err;
1762 }
1763
1764 void xenvif_stop_queue(struct xenvif *vif)
1765 {
1766 if (!vif->can_queue)
1767 return;
1768
1769 netif_stop_queue(vif->dev);
1770 }
1771
1772 static void xenvif_start_queue(struct xenvif *vif)
1773 {
1774 if (xenvif_schedulable(vif))
1775 netif_wake_queue(vif->dev);
1776 }
1777
1778 int xenvif_kthread_guest_rx(void *data)
1779 {
1780 struct xenvif *vif = data;
1781 struct sk_buff *skb;
1782
1783 while (!kthread_should_stop()) {
1784 wait_event_interruptible(vif->wq,
1785 rx_work_todo(vif) ||
1786 kthread_should_stop());
1787 if (kthread_should_stop())
1788 break;
1789
1790 if (vif->rx_queue_purge) {
1791 skb_queue_purge(&vif->rx_queue);
1792 vif->rx_queue_purge = false;
1793 }
1794
1795 if (!skb_queue_empty(&vif->rx_queue))
1796 xenvif_rx_action(vif);
1797
1798 if (skb_queue_empty(&vif->rx_queue) &&
1799 netif_queue_stopped(vif->dev)) {
1800 del_timer_sync(&vif->wake_queue);
1801 xenvif_start_queue(vif);
1802 }
1803
1804 cond_resched();
1805 }
1806
1807 /* Bin any remaining skbs */
1808 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL)
1809 dev_kfree_skb(skb);
1810
1811 return 0;
1812 }
1813
1814 int xenvif_dealloc_kthread(void *data)
1815 {
1816 struct xenvif *vif = data;
1817
1818 while (!kthread_should_stop()) {
1819 wait_event_interruptible(vif->dealloc_wq,
1820 tx_dealloc_work_todo(vif) ||
1821 kthread_should_stop());
1822 if (kthread_should_stop())
1823 break;
1824
1825 xenvif_tx_dealloc_action(vif);
1826 cond_resched();
1827 }
1828
1829 /* Unmap anything remaining*/
1830 if (tx_dealloc_work_todo(vif))
1831 xenvif_tx_dealloc_action(vif);
1832
1833 return 0;
1834 }
1835
1836 static int __init netback_init(void)
1837 {
1838 int rc = 0;
1839
1840 if (!xen_domain())
1841 return -ENODEV;
1842
1843 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1844 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1845 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1846 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1847 }
1848
1849 rc = xenvif_xenbus_init();
1850 if (rc)
1851 goto failed_init;
1852
1853 rx_drain_timeout_jiffies = msecs_to_jiffies(rx_drain_timeout_msecs);
1854
1855 return 0;
1856
1857 failed_init:
1858 return rc;
1859 }
1860
1861 module_init(netback_init);
1862
1863 static void __exit netback_fini(void)
1864 {
1865 xenvif_xenbus_fini();
1866 }
1867 module_exit(netback_fini);
1868
1869 MODULE_LICENSE("Dual BSD/GPL");
1870 MODULE_ALIAS("xen-backend:vif");
Linux kernel - Deliverables
This page took 0.321782 seconds and 5 git commands to generate.