Loading...
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#include <xen/page.h>
48
49#include <asm/xen/hypercall.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 */
55bool separate_tx_rx_irq = true;
56module_param(separate_tx_rx_irq, bool, 0644);
57
58/* The time that packets can stay on the guest Rx internal queue
59 * before they are dropped.
60 */
61unsigned int rx_drain_timeout_msecs = 10000;
62module_param(rx_drain_timeout_msecs, uint, 0444);
63
64/* The length of time before the frontend is considered unresponsive
65 * because it isn't providing Rx slots.
66 */
67unsigned int rx_stall_timeout_msecs = 60000;
68module_param(rx_stall_timeout_msecs, uint, 0444);
69
70#define MAX_QUEUES_DEFAULT 8
71unsigned int xenvif_max_queues;
72module_param_named(max_queues, xenvif_max_queues, uint, 0644);
73MODULE_PARM_DESC(max_queues,
74 "Maximum number of queues per virtual interface");
75
76/*
77 * This is the maximum slots a skb can have. If a guest sends a skb
78 * which exceeds this limit it is considered malicious.
79 */
80#define FATAL_SKB_SLOTS_DEFAULT 20
81static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
82module_param(fatal_skb_slots, uint, 0444);
83
84/* The amount to copy out of the first guest Tx slot into the skb's
85 * linear area. If the first slot has more data, it will be mapped
86 * and put into the first frag.
87 *
88 * This is sized to avoid pulling headers from the frags for most
89 * TCP/IP packets.
90 */
91#define XEN_NETBACK_TX_COPY_LEN 128
92
93/* This is the maximum number of flows in the hash cache. */
94#define XENVIF_HASH_CACHE_SIZE_DEFAULT 64
95unsigned int xenvif_hash_cache_size = XENVIF_HASH_CACHE_SIZE_DEFAULT;
96module_param_named(hash_cache_size, xenvif_hash_cache_size, uint, 0644);
97MODULE_PARM_DESC(hash_cache_size, "Number of flows in the hash cache");
98
99/* The module parameter tells that we have to put data
100 * for xen-netfront with the XDP_PACKET_HEADROOM offset
101 * needed for XDP processing
102 */
103bool provides_xdp_headroom = true;
104module_param(provides_xdp_headroom, bool, 0644);
105
106static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
107 s8 status);
108
109static void make_tx_response(struct xenvif_queue *queue,
110 const struct xen_netif_tx_request *txp,
111 unsigned int extra_count,
112 s8 status);
113
114static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx);
115
116static inline int tx_work_todo(struct xenvif_queue *queue);
117
118static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
119 u16 idx)
120{
121 return page_to_pfn(queue->mmap_pages[idx]);
122}
123
124static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
125 u16 idx)
126{
127 return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
128}
129
130#define callback_param(vif, pending_idx) \
131 (vif->pending_tx_info[pending_idx].callback_struct)
132
133/* Find the containing VIF's structure from a pointer in pending_tx_info array
134 */
135static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info_msgzc *ubuf)
136{
137 u16 pending_idx = ubuf->desc;
138 struct pending_tx_info *temp =
139 container_of(ubuf, struct pending_tx_info, callback_struct);
140 return container_of(temp - pending_idx,
141 struct xenvif_queue,
142 pending_tx_info[0]);
143}
144
145static u16 frag_get_pending_idx(skb_frag_t *frag)
146{
147 return (u16)skb_frag_off(frag);
148}
149
150static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
151{
152 skb_frag_off_set(frag, pending_idx);
153}
154
155static inline pending_ring_idx_t pending_index(unsigned i)
156{
157 return i & (MAX_PENDING_REQS-1);
158}
159
160void xenvif_kick_thread(struct xenvif_queue *queue)
161{
162 wake_up(&queue->wq);
163}
164
165void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
166{
167 int more_to_do;
168
169 RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
170
171 if (more_to_do)
172 napi_schedule(&queue->napi);
173 else if (atomic_fetch_andnot(NETBK_TX_EOI | NETBK_COMMON_EOI,
174 &queue->eoi_pending) &
175 (NETBK_TX_EOI | NETBK_COMMON_EOI))
176 xen_irq_lateeoi(queue->tx_irq, 0);
177}
178
179static void tx_add_credit(struct xenvif_queue *queue)
180{
181 unsigned long max_burst, max_credit;
182
183 /*
184 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
185 * Otherwise the interface can seize up due to insufficient credit.
186 */
187 max_burst = max(131072UL, queue->credit_bytes);
188
189 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
190 max_credit = queue->remaining_credit + queue->credit_bytes;
191 if (max_credit < queue->remaining_credit)
192 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
193
194 queue->remaining_credit = min(max_credit, max_burst);
195 queue->rate_limited = false;
196}
197
198void xenvif_tx_credit_callback(struct timer_list *t)
199{
200 struct xenvif_queue *queue = from_timer(queue, t, credit_timeout);
201 tx_add_credit(queue);
202 xenvif_napi_schedule_or_enable_events(queue);
203}
204
205static void xenvif_tx_err(struct xenvif_queue *queue,
206 struct xen_netif_tx_request *txp,
207 unsigned int extra_count, RING_IDX end)
208{
209 RING_IDX cons = queue->tx.req_cons;
210
211 do {
212 make_tx_response(queue, txp, extra_count, XEN_NETIF_RSP_ERROR);
213 if (cons == end)
214 break;
215 RING_COPY_REQUEST(&queue->tx, cons++, txp);
216 extra_count = 0; /* only the first frag can have extras */
217 } while (1);
218 queue->tx.req_cons = cons;
219}
220
221static void xenvif_fatal_tx_err(struct xenvif *vif)
222{
223 netdev_err(vif->dev, "fatal error; disabling device\n");
224 vif->disabled = true;
225 /* Disable the vif from queue 0's kthread */
226 if (vif->num_queues)
227 xenvif_kick_thread(&vif->queues[0]);
228}
229
230static int xenvif_count_requests(struct xenvif_queue *queue,
231 struct xen_netif_tx_request *first,
232 unsigned int extra_count,
233 struct xen_netif_tx_request *txp,
234 int work_to_do)
235{
236 RING_IDX cons = queue->tx.req_cons;
237 int slots = 0;
238 int drop_err = 0;
239 int more_data;
240
241 if (!(first->flags & XEN_NETTXF_more_data))
242 return 0;
243
244 do {
245 struct xen_netif_tx_request dropped_tx = { 0 };
246
247 if (slots >= work_to_do) {
248 netdev_err(queue->vif->dev,
249 "Asked for %d slots but exceeds this limit\n",
250 work_to_do);
251 xenvif_fatal_tx_err(queue->vif);
252 return -ENODATA;
253 }
254
255 /* This guest is really using too many slots and
256 * considered malicious.
257 */
258 if (unlikely(slots >= fatal_skb_slots)) {
259 netdev_err(queue->vif->dev,
260 "Malicious frontend using %d slots, threshold %u\n",
261 slots, fatal_skb_slots);
262 xenvif_fatal_tx_err(queue->vif);
263 return -E2BIG;
264 }
265
266 /* Xen network protocol had implicit dependency on
267 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
268 * the historical MAX_SKB_FRAGS value 18 to honor the
269 * same behavior as before. Any packet using more than
270 * 18 slots but less than fatal_skb_slots slots is
271 * dropped
272 */
273 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
274 if (net_ratelimit())
275 netdev_dbg(queue->vif->dev,
276 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
277 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
278 drop_err = -E2BIG;
279 }
280
281 if (drop_err)
282 txp = &dropped_tx;
283
284 RING_COPY_REQUEST(&queue->tx, cons + slots, txp);
285
286 /* If the guest submitted a frame >= 64 KiB then
287 * first->size overflowed and following slots will
288 * appear to be larger than the frame.
289 *
290 * This cannot be fatal error as there are buggy
291 * frontends that do this.
292 *
293 * Consume all slots and drop the packet.
294 */
295 if (!drop_err && txp->size > first->size) {
296 if (net_ratelimit())
297 netdev_dbg(queue->vif->dev,
298 "Invalid tx request, slot size %u > remaining size %u\n",
299 txp->size, first->size);
300 drop_err = -EIO;
301 }
302
303 first->size -= txp->size;
304 slots++;
305
306 if (unlikely((txp->offset + txp->size) > XEN_PAGE_SIZE)) {
307 netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %u, size: %u\n",
308 txp->offset, txp->size);
309 xenvif_fatal_tx_err(queue->vif);
310 return -EINVAL;
311 }
312
313 more_data = txp->flags & XEN_NETTXF_more_data;
314
315 if (!drop_err)
316 txp++;
317
318 } while (more_data);
319
320 if (drop_err) {
321 xenvif_tx_err(queue, first, extra_count, cons + slots);
322 return drop_err;
323 }
324
325 return slots;
326}
327
328
329struct xenvif_tx_cb {
330 u16 copy_pending_idx[XEN_NETBK_LEGACY_SLOTS_MAX + 1];
331 u8 copy_count;
332 u32 split_mask;
333};
334
335#define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
336#define copy_pending_idx(skb, i) (XENVIF_TX_CB(skb)->copy_pending_idx[i])
337#define copy_count(skb) (XENVIF_TX_CB(skb)->copy_count)
338
339static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
340 u16 pending_idx,
341 struct xen_netif_tx_request *txp,
342 unsigned int extra_count,
343 struct gnttab_map_grant_ref *mop)
344{
345 queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
346 gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
347 GNTMAP_host_map | GNTMAP_readonly,
348 txp->gref, queue->vif->domid);
349
350 memcpy(&queue->pending_tx_info[pending_idx].req, txp,
351 sizeof(*txp));
352 queue->pending_tx_info[pending_idx].extra_count = extra_count;
353}
354
355static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
356{
357 struct sk_buff *skb =
358 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
359 GFP_ATOMIC | __GFP_NOWARN);
360
361 BUILD_BUG_ON(sizeof(*XENVIF_TX_CB(skb)) > sizeof(skb->cb));
362 if (unlikely(skb == NULL))
363 return NULL;
364
365 /* Packets passed to netif_rx() must have some headroom. */
366 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
367
368 /* Initialize it here to avoid later surprises */
369 skb_shinfo(skb)->destructor_arg = NULL;
370
371 return skb;
372}
373
374static void xenvif_get_requests(struct xenvif_queue *queue,
375 struct sk_buff *skb,
376 struct xen_netif_tx_request *first,
377 struct xen_netif_tx_request *txfrags,
378 unsigned *copy_ops,
379 unsigned *map_ops,
380 unsigned int frag_overflow,
381 struct sk_buff *nskb,
382 unsigned int extra_count,
383 unsigned int data_len)
384{
385 struct skb_shared_info *shinfo = skb_shinfo(skb);
386 skb_frag_t *frags = shinfo->frags;
387 u16 pending_idx;
388 pending_ring_idx_t index;
389 unsigned int nr_slots;
390 struct gnttab_copy *cop = queue->tx_copy_ops + *copy_ops;
391 struct gnttab_map_grant_ref *gop = queue->tx_map_ops + *map_ops;
392 struct xen_netif_tx_request *txp = first;
393
394 nr_slots = shinfo->nr_frags + frag_overflow + 1;
395
396 copy_count(skb) = 0;
397 XENVIF_TX_CB(skb)->split_mask = 0;
398
399 /* Create copy ops for exactly data_len bytes into the skb head. */
400 __skb_put(skb, data_len);
401 while (data_len > 0) {
402 int amount = data_len > txp->size ? txp->size : data_len;
403 bool split = false;
404
405 cop->source.u.ref = txp->gref;
406 cop->source.domid = queue->vif->domid;
407 cop->source.offset = txp->offset;
408
409 cop->dest.domid = DOMID_SELF;
410 cop->dest.offset = (offset_in_page(skb->data +
411 skb_headlen(skb) -
412 data_len)) & ~XEN_PAGE_MASK;
413 cop->dest.u.gmfn = virt_to_gfn(skb->data + skb_headlen(skb)
414 - data_len);
415
416 /* Don't cross local page boundary! */
417 if (cop->dest.offset + amount > XEN_PAGE_SIZE) {
418 amount = XEN_PAGE_SIZE - cop->dest.offset;
419 XENVIF_TX_CB(skb)->split_mask |= 1U << copy_count(skb);
420 split = true;
421 }
422
423 cop->len = amount;
424 cop->flags = GNTCOPY_source_gref;
425
426 index = pending_index(queue->pending_cons);
427 pending_idx = queue->pending_ring[index];
428 callback_param(queue, pending_idx).ctx = NULL;
429 copy_pending_idx(skb, copy_count(skb)) = pending_idx;
430 if (!split)
431 copy_count(skb)++;
432
433 cop++;
434 data_len -= amount;
435
436 if (amount == txp->size) {
437 /* The copy op covered the full tx_request */
438
439 memcpy(&queue->pending_tx_info[pending_idx].req,
440 txp, sizeof(*txp));
441 queue->pending_tx_info[pending_idx].extra_count =
442 (txp == first) ? extra_count : 0;
443
444 if (txp == first)
445 txp = txfrags;
446 else
447 txp++;
448 queue->pending_cons++;
449 nr_slots--;
450 } else {
451 /* The copy op partially covered the tx_request.
452 * The remainder will be mapped or copied in the next
453 * iteration.
454 */
455 txp->offset += amount;
456 txp->size -= amount;
457 }
458 }
459
460 for (shinfo->nr_frags = 0; nr_slots > 0 && shinfo->nr_frags < MAX_SKB_FRAGS;
461 nr_slots--) {
462 if (unlikely(!txp->size)) {
463 make_tx_response(queue, txp, 0, XEN_NETIF_RSP_OKAY);
464 ++txp;
465 continue;
466 }
467
468 index = pending_index(queue->pending_cons++);
469 pending_idx = queue->pending_ring[index];
470 xenvif_tx_create_map_op(queue, pending_idx, txp,
471 txp == first ? extra_count : 0, gop);
472 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
473 ++shinfo->nr_frags;
474 ++gop;
475
476 if (txp == first)
477 txp = txfrags;
478 else
479 txp++;
480 }
481
482 if (nr_slots > 0) {
483
484 shinfo = skb_shinfo(nskb);
485 frags = shinfo->frags;
486
487 for (shinfo->nr_frags = 0; shinfo->nr_frags < nr_slots; ++txp) {
488 if (unlikely(!txp->size)) {
489 make_tx_response(queue, txp, 0,
490 XEN_NETIF_RSP_OKAY);
491 continue;
492 }
493
494 index = pending_index(queue->pending_cons++);
495 pending_idx = queue->pending_ring[index];
496 xenvif_tx_create_map_op(queue, pending_idx, txp, 0,
497 gop);
498 frag_set_pending_idx(&frags[shinfo->nr_frags],
499 pending_idx);
500 ++shinfo->nr_frags;
501 ++gop;
502 }
503
504 if (shinfo->nr_frags) {
505 skb_shinfo(skb)->frag_list = nskb;
506 nskb = NULL;
507 }
508 }
509
510 if (nskb) {
511 /* A frag_list skb was allocated but it is no longer needed
512 * because enough slots were converted to copy ops above or some
513 * were empty.
514 */
515 kfree_skb(nskb);
516 }
517
518 (*copy_ops) = cop - queue->tx_copy_ops;
519 (*map_ops) = gop - queue->tx_map_ops;
520}
521
522static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
523 u16 pending_idx,
524 grant_handle_t handle)
525{
526 if (unlikely(queue->grant_tx_handle[pending_idx] !=
527 NETBACK_INVALID_HANDLE)) {
528 netdev_err(queue->vif->dev,
529 "Trying to overwrite active handle! pending_idx: 0x%x\n",
530 pending_idx);
531 BUG();
532 }
533 queue->grant_tx_handle[pending_idx] = handle;
534}
535
536static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
537 u16 pending_idx)
538{
539 if (unlikely(queue->grant_tx_handle[pending_idx] ==
540 NETBACK_INVALID_HANDLE)) {
541 netdev_err(queue->vif->dev,
542 "Trying to unmap invalid handle! pending_idx: 0x%x\n",
543 pending_idx);
544 BUG();
545 }
546 queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
547}
548
549static int xenvif_tx_check_gop(struct xenvif_queue *queue,
550 struct sk_buff *skb,
551 struct gnttab_map_grant_ref **gopp_map,
552 struct gnttab_copy **gopp_copy)
553{
554 struct gnttab_map_grant_ref *gop_map = *gopp_map;
555 u16 pending_idx;
556 /* This always points to the shinfo of the skb being checked, which
557 * could be either the first or the one on the frag_list
558 */
559 struct skb_shared_info *shinfo = skb_shinfo(skb);
560 /* If this is non-NULL, we are currently checking the frag_list skb, and
561 * this points to the shinfo of the first one
562 */
563 struct skb_shared_info *first_shinfo = NULL;
564 int nr_frags = shinfo->nr_frags;
565 const bool sharedslot = nr_frags &&
566 frag_get_pending_idx(&shinfo->frags[0]) ==
567 copy_pending_idx(skb, copy_count(skb) - 1);
568 int i, err = 0;
569
570 for (i = 0; i < copy_count(skb); i++) {
571 int newerr;
572
573 /* Check status of header. */
574 pending_idx = copy_pending_idx(skb, i);
575
576 newerr = (*gopp_copy)->status;
577
578 /* Split copies need to be handled together. */
579 if (XENVIF_TX_CB(skb)->split_mask & (1U << i)) {
580 (*gopp_copy)++;
581 if (!newerr)
582 newerr = (*gopp_copy)->status;
583 }
584 if (likely(!newerr)) {
585 /* The first frag might still have this slot mapped */
586 if (i < copy_count(skb) - 1 || !sharedslot)
587 xenvif_idx_release(queue, pending_idx,
588 XEN_NETIF_RSP_OKAY);
589 } else {
590 err = newerr;
591 if (net_ratelimit())
592 netdev_dbg(queue->vif->dev,
593 "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
594 (*gopp_copy)->status,
595 pending_idx,
596 (*gopp_copy)->source.u.ref);
597 /* The first frag might still have this slot mapped */
598 if (i < copy_count(skb) - 1 || !sharedslot)
599 xenvif_idx_release(queue, pending_idx,
600 XEN_NETIF_RSP_ERROR);
601 }
602 (*gopp_copy)++;
603 }
604
605check_frags:
606 for (i = 0; i < nr_frags; i++, gop_map++) {
607 int j, newerr;
608
609 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
610
611 /* Check error status: if okay then remember grant handle. */
612 newerr = gop_map->status;
613
614 if (likely(!newerr)) {
615 xenvif_grant_handle_set(queue,
616 pending_idx,
617 gop_map->handle);
618 /* Had a previous error? Invalidate this fragment. */
619 if (unlikely(err)) {
620 xenvif_idx_unmap(queue, pending_idx);
621 /* If the mapping of the first frag was OK, but
622 * the header's copy failed, and they are
623 * sharing a slot, send an error
624 */
625 if (i == 0 && !first_shinfo && sharedslot)
626 xenvif_idx_release(queue, pending_idx,
627 XEN_NETIF_RSP_ERROR);
628 else
629 xenvif_idx_release(queue, pending_idx,
630 XEN_NETIF_RSP_OKAY);
631 }
632 continue;
633 }
634
635 /* Error on this fragment: respond to client with an error. */
636 if (net_ratelimit())
637 netdev_dbg(queue->vif->dev,
638 "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
639 i,
640 gop_map->status,
641 pending_idx,
642 gop_map->ref);
643
644 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
645
646 /* Not the first error? Preceding frags already invalidated. */
647 if (err)
648 continue;
649
650 /* Invalidate preceding fragments of this skb. */
651 for (j = 0; j < i; j++) {
652 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
653 xenvif_idx_unmap(queue, pending_idx);
654 xenvif_idx_release(queue, pending_idx,
655 XEN_NETIF_RSP_OKAY);
656 }
657
658 /* And if we found the error while checking the frag_list, unmap
659 * the first skb's frags
660 */
661 if (first_shinfo) {
662 for (j = 0; j < first_shinfo->nr_frags; j++) {
663 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
664 xenvif_idx_unmap(queue, pending_idx);
665 xenvif_idx_release(queue, pending_idx,
666 XEN_NETIF_RSP_OKAY);
667 }
668 }
669
670 /* Remember the error: invalidate all subsequent fragments. */
671 err = newerr;
672 }
673
674 if (skb_has_frag_list(skb) && !first_shinfo) {
675 first_shinfo = shinfo;
676 shinfo = skb_shinfo(shinfo->frag_list);
677 nr_frags = shinfo->nr_frags;
678
679 goto check_frags;
680 }
681
682 *gopp_map = gop_map;
683 return err;
684}
685
686static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
687{
688 struct skb_shared_info *shinfo = skb_shinfo(skb);
689 int nr_frags = shinfo->nr_frags;
690 int i;
691 u16 prev_pending_idx = INVALID_PENDING_IDX;
692
693 for (i = 0; i < nr_frags; i++) {
694 skb_frag_t *frag = shinfo->frags + i;
695 struct xen_netif_tx_request *txp;
696 struct page *page;
697 u16 pending_idx;
698
699 pending_idx = frag_get_pending_idx(frag);
700
701 /* If this is not the first frag, chain it to the previous*/
702 if (prev_pending_idx == INVALID_PENDING_IDX)
703 skb_shinfo(skb)->destructor_arg =
704 &callback_param(queue, pending_idx);
705 else
706 callback_param(queue, prev_pending_idx).ctx =
707 &callback_param(queue, pending_idx);
708
709 callback_param(queue, pending_idx).ctx = NULL;
710 prev_pending_idx = pending_idx;
711
712 txp = &queue->pending_tx_info[pending_idx].req;
713 page = virt_to_page((void *)idx_to_kaddr(queue, pending_idx));
714 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
715 skb->len += txp->size;
716 skb->data_len += txp->size;
717 skb->truesize += txp->size;
718
719 /* Take an extra reference to offset network stack's put_page */
720 get_page(queue->mmap_pages[pending_idx]);
721 }
722}
723
724static int xenvif_get_extras(struct xenvif_queue *queue,
725 struct xen_netif_extra_info *extras,
726 unsigned int *extra_count,
727 int work_to_do)
728{
729 struct xen_netif_extra_info extra;
730 RING_IDX cons = queue->tx.req_cons;
731
732 do {
733 if (unlikely(work_to_do-- <= 0)) {
734 netdev_err(queue->vif->dev, "Missing extra info\n");
735 xenvif_fatal_tx_err(queue->vif);
736 return -EBADR;
737 }
738
739 RING_COPY_REQUEST(&queue->tx, cons, &extra);
740
741 queue->tx.req_cons = ++cons;
742 (*extra_count)++;
743
744 if (unlikely(!extra.type ||
745 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
746 netdev_err(queue->vif->dev,
747 "Invalid extra type: %d\n", extra.type);
748 xenvif_fatal_tx_err(queue->vif);
749 return -EINVAL;
750 }
751
752 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
753 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
754
755 return work_to_do;
756}
757
758static int xenvif_set_skb_gso(struct xenvif *vif,
759 struct sk_buff *skb,
760 struct xen_netif_extra_info *gso)
761{
762 if (!gso->u.gso.size) {
763 netdev_err(vif->dev, "GSO size must not be zero.\n");
764 xenvif_fatal_tx_err(vif);
765 return -EINVAL;
766 }
767
768 switch (gso->u.gso.type) {
769 case XEN_NETIF_GSO_TYPE_TCPV4:
770 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
771 break;
772 case XEN_NETIF_GSO_TYPE_TCPV6:
773 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
774 break;
775 default:
776 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
777 xenvif_fatal_tx_err(vif);
778 return -EINVAL;
779 }
780
781 skb_shinfo(skb)->gso_size = gso->u.gso.size;
782 /* gso_segs will be calculated later */
783
784 return 0;
785}
786
787static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
788{
789 bool recalculate_partial_csum = false;
790
791 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
792 * peers can fail to set NETRXF_csum_blank when sending a GSO
793 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
794 * recalculate the partial checksum.
795 */
796 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
797 queue->stats.rx_gso_checksum_fixup++;
798 skb->ip_summed = CHECKSUM_PARTIAL;
799 recalculate_partial_csum = true;
800 }
801
802 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
803 if (skb->ip_summed != CHECKSUM_PARTIAL)
804 return 0;
805
806 return skb_checksum_setup(skb, recalculate_partial_csum);
807}
808
809static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
810{
811 u64 now = get_jiffies_64();
812 u64 next_credit = queue->credit_window_start +
813 msecs_to_jiffies(queue->credit_usec / 1000);
814
815 /* Timer could already be pending in rare cases. */
816 if (timer_pending(&queue->credit_timeout)) {
817 queue->rate_limited = true;
818 return true;
819 }
820
821 /* Passed the point where we can replenish credit? */
822 if (time_after_eq64(now, next_credit)) {
823 queue->credit_window_start = now;
824 tx_add_credit(queue);
825 }
826
827 /* Still too big to send right now? Set a callback. */
828 if (size > queue->remaining_credit) {
829 mod_timer(&queue->credit_timeout,
830 next_credit);
831 queue->credit_window_start = next_credit;
832 queue->rate_limited = true;
833
834 return true;
835 }
836
837 return false;
838}
839
840/* No locking is required in xenvif_mcast_add/del() as they are
841 * only ever invoked from NAPI poll. An RCU list is used because
842 * xenvif_mcast_match() is called asynchronously, during start_xmit.
843 */
844
845static int xenvif_mcast_add(struct xenvif *vif, const u8 *addr)
846{
847 struct xenvif_mcast_addr *mcast;
848
849 if (vif->fe_mcast_count == XEN_NETBK_MCAST_MAX) {
850 if (net_ratelimit())
851 netdev_err(vif->dev,
852 "Too many multicast addresses\n");
853 return -ENOSPC;
854 }
855
856 mcast = kzalloc(sizeof(*mcast), GFP_ATOMIC);
857 if (!mcast)
858 return -ENOMEM;
859
860 ether_addr_copy(mcast->addr, addr);
861 list_add_tail_rcu(&mcast->entry, &vif->fe_mcast_addr);
862 vif->fe_mcast_count++;
863
864 return 0;
865}
866
867static void xenvif_mcast_del(struct xenvif *vif, const u8 *addr)
868{
869 struct xenvif_mcast_addr *mcast;
870
871 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
872 if (ether_addr_equal(addr, mcast->addr)) {
873 --vif->fe_mcast_count;
874 list_del_rcu(&mcast->entry);
875 kfree_rcu(mcast, rcu);
876 break;
877 }
878 }
879}
880
881bool xenvif_mcast_match(struct xenvif *vif, const u8 *addr)
882{
883 struct xenvif_mcast_addr *mcast;
884
885 rcu_read_lock();
886 list_for_each_entry_rcu(mcast, &vif->fe_mcast_addr, entry) {
887 if (ether_addr_equal(addr, mcast->addr)) {
888 rcu_read_unlock();
889 return true;
890 }
891 }
892 rcu_read_unlock();
893
894 return false;
895}
896
897void xenvif_mcast_addr_list_free(struct xenvif *vif)
898{
899 /* No need for locking or RCU here. NAPI poll and TX queue
900 * are stopped.
901 */
902 while (!list_empty(&vif->fe_mcast_addr)) {
903 struct xenvif_mcast_addr *mcast;
904
905 mcast = list_first_entry(&vif->fe_mcast_addr,
906 struct xenvif_mcast_addr,
907 entry);
908 --vif->fe_mcast_count;
909 list_del(&mcast->entry);
910 kfree(mcast);
911 }
912}
913
914static void xenvif_tx_build_gops(struct xenvif_queue *queue,
915 int budget,
916 unsigned *copy_ops,
917 unsigned *map_ops)
918{
919 struct sk_buff *skb, *nskb;
920 int ret;
921 unsigned int frag_overflow;
922
923 while (skb_queue_len(&queue->tx_queue) < budget) {
924 struct xen_netif_tx_request txreq;
925 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
926 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
927 unsigned int extra_count;
928 RING_IDX idx;
929 int work_to_do;
930 unsigned int data_len;
931
932 if (queue->tx.sring->req_prod - queue->tx.req_cons >
933 XEN_NETIF_TX_RING_SIZE) {
934 netdev_err(queue->vif->dev,
935 "Impossible number of requests. "
936 "req_prod %d, req_cons %d, size %ld\n",
937 queue->tx.sring->req_prod, queue->tx.req_cons,
938 XEN_NETIF_TX_RING_SIZE);
939 xenvif_fatal_tx_err(queue->vif);
940 break;
941 }
942
943 work_to_do = XEN_RING_NR_UNCONSUMED_REQUESTS(&queue->tx);
944 if (!work_to_do)
945 break;
946
947 idx = queue->tx.req_cons;
948 rmb(); /* Ensure that we see the request before we copy it. */
949 RING_COPY_REQUEST(&queue->tx, idx, &txreq);
950
951 /* Credit-based scheduling. */
952 if (txreq.size > queue->remaining_credit &&
953 tx_credit_exceeded(queue, txreq.size))
954 break;
955
956 queue->remaining_credit -= txreq.size;
957
958 work_to_do--;
959 queue->tx.req_cons = ++idx;
960
961 memset(extras, 0, sizeof(extras));
962 extra_count = 0;
963 if (txreq.flags & XEN_NETTXF_extra_info) {
964 work_to_do = xenvif_get_extras(queue, extras,
965 &extra_count,
966 work_to_do);
967 idx = queue->tx.req_cons;
968 if (unlikely(work_to_do < 0))
969 break;
970 }
971
972 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1].type) {
973 struct xen_netif_extra_info *extra;
974
975 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_ADD - 1];
976 ret = xenvif_mcast_add(queue->vif, extra->u.mcast.addr);
977
978 make_tx_response(queue, &txreq, extra_count,
979 (ret == 0) ?
980 XEN_NETIF_RSP_OKAY :
981 XEN_NETIF_RSP_ERROR);
982 continue;
983 }
984
985 if (extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1].type) {
986 struct xen_netif_extra_info *extra;
987
988 extra = &extras[XEN_NETIF_EXTRA_TYPE_MCAST_DEL - 1];
989 xenvif_mcast_del(queue->vif, extra->u.mcast.addr);
990
991 make_tx_response(queue, &txreq, extra_count,
992 XEN_NETIF_RSP_OKAY);
993 continue;
994 }
995
996 data_len = (txreq.size > XEN_NETBACK_TX_COPY_LEN) ?
997 XEN_NETBACK_TX_COPY_LEN : txreq.size;
998
999 ret = xenvif_count_requests(queue, &txreq, extra_count,
1000 txfrags, work_to_do);
1001
1002 if (unlikely(ret < 0))
1003 break;
1004
1005 idx += ret;
1006
1007 if (unlikely(txreq.size < ETH_HLEN)) {
1008 netdev_dbg(queue->vif->dev,
1009 "Bad packet size: %d\n", txreq.size);
1010 xenvif_tx_err(queue, &txreq, extra_count, idx);
1011 break;
1012 }
1013
1014 /* No crossing a page as the payload mustn't fragment. */
1015 if (unlikely((txreq.offset + txreq.size) > XEN_PAGE_SIZE)) {
1016 netdev_err(queue->vif->dev, "Cross page boundary, txreq.offset: %u, size: %u\n",
1017 txreq.offset, txreq.size);
1018 xenvif_fatal_tx_err(queue->vif);
1019 break;
1020 }
1021
1022 if (ret >= XEN_NETBK_LEGACY_SLOTS_MAX - 1 && data_len < txreq.size)
1023 data_len = txreq.size;
1024
1025 skb = xenvif_alloc_skb(data_len);
1026 if (unlikely(skb == NULL)) {
1027 netdev_dbg(queue->vif->dev,
1028 "Can't allocate a skb in start_xmit.\n");
1029 xenvif_tx_err(queue, &txreq, extra_count, idx);
1030 break;
1031 }
1032
1033 skb_shinfo(skb)->nr_frags = ret;
1034 /* At this point shinfo->nr_frags is in fact the number of
1035 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
1036 */
1037 frag_overflow = 0;
1038 nskb = NULL;
1039 if (skb_shinfo(skb)->nr_frags > MAX_SKB_FRAGS) {
1040 frag_overflow = skb_shinfo(skb)->nr_frags - MAX_SKB_FRAGS;
1041 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
1042 skb_shinfo(skb)->nr_frags = MAX_SKB_FRAGS;
1043 nskb = xenvif_alloc_skb(0);
1044 if (unlikely(nskb == NULL)) {
1045 skb_shinfo(skb)->nr_frags = 0;
1046 kfree_skb(skb);
1047 xenvif_tx_err(queue, &txreq, extra_count, idx);
1048 if (net_ratelimit())
1049 netdev_err(queue->vif->dev,
1050 "Can't allocate the frag_list skb.\n");
1051 break;
1052 }
1053 }
1054
1055 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1056 struct xen_netif_extra_info *gso;
1057 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1058
1059 if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1060 /* Failure in xenvif_set_skb_gso is fatal. */
1061 skb_shinfo(skb)->nr_frags = 0;
1062 kfree_skb(skb);
1063 kfree_skb(nskb);
1064 break;
1065 }
1066 }
1067
1068 if (extras[XEN_NETIF_EXTRA_TYPE_HASH - 1].type) {
1069 struct xen_netif_extra_info *extra;
1070 enum pkt_hash_types type = PKT_HASH_TYPE_NONE;
1071
1072 extra = &extras[XEN_NETIF_EXTRA_TYPE_HASH - 1];
1073
1074 switch (extra->u.hash.type) {
1075 case _XEN_NETIF_CTRL_HASH_TYPE_IPV4:
1076 case _XEN_NETIF_CTRL_HASH_TYPE_IPV6:
1077 type = PKT_HASH_TYPE_L3;
1078 break;
1079
1080 case _XEN_NETIF_CTRL_HASH_TYPE_IPV4_TCP:
1081 case _XEN_NETIF_CTRL_HASH_TYPE_IPV6_TCP:
1082 type = PKT_HASH_TYPE_L4;
1083 break;
1084
1085 default:
1086 break;
1087 }
1088
1089 if (type != PKT_HASH_TYPE_NONE)
1090 skb_set_hash(skb,
1091 *(u32 *)extra->u.hash.value,
1092 type);
1093 }
1094
1095 xenvif_get_requests(queue, skb, &txreq, txfrags, copy_ops,
1096 map_ops, frag_overflow, nskb, extra_count,
1097 data_len);
1098
1099 __skb_queue_tail(&queue->tx_queue, skb);
1100
1101 queue->tx.req_cons = idx;
1102 }
1103
1104 return;
1105}
1106
1107/* Consolidate skb with a frag_list into a brand new one with local pages on
1108 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1109 */
1110static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1111{
1112 unsigned int offset = skb_headlen(skb);
1113 skb_frag_t frags[MAX_SKB_FRAGS];
1114 int i, f;
1115 struct ubuf_info *uarg;
1116 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1117
1118 queue->stats.tx_zerocopy_sent += 2;
1119 queue->stats.tx_frag_overflow++;
1120
1121 xenvif_fill_frags(queue, nskb);
1122 /* Subtract frags size, we will correct it later */
1123 skb->truesize -= skb->data_len;
1124 skb->len += nskb->len;
1125 skb->data_len += nskb->len;
1126
1127 /* create a brand new frags array and coalesce there */
1128 for (i = 0; offset < skb->len; i++) {
1129 struct page *page;
1130 unsigned int len;
1131
1132 BUG_ON(i >= MAX_SKB_FRAGS);
1133 page = alloc_page(GFP_ATOMIC);
1134 if (!page) {
1135 int j;
1136 skb->truesize += skb->data_len;
1137 for (j = 0; j < i; j++)
1138 put_page(skb_frag_page(&frags[j]));
1139 return -ENOMEM;
1140 }
1141
1142 if (offset + PAGE_SIZE < skb->len)
1143 len = PAGE_SIZE;
1144 else
1145 len = skb->len - offset;
1146 if (skb_copy_bits(skb, offset, page_address(page), len))
1147 BUG();
1148
1149 offset += len;
1150 skb_frag_fill_page_desc(&frags[i], page, 0, len);
1151 }
1152
1153 /* Release all the original (foreign) frags. */
1154 for (f = 0; f < skb_shinfo(skb)->nr_frags; f++)
1155 skb_frag_unref(skb, f);
1156 uarg = skb_shinfo(skb)->destructor_arg;
1157 /* increase inflight counter to offset decrement in callback */
1158 atomic_inc(&queue->inflight_packets);
1159 uarg->callback(NULL, uarg, true);
1160 skb_shinfo(skb)->destructor_arg = NULL;
1161
1162 /* Fill the skb with the new (local) frags. */
1163 memcpy(skb_shinfo(skb)->frags, frags, i * sizeof(skb_frag_t));
1164 skb_shinfo(skb)->nr_frags = i;
1165 skb->truesize += i * PAGE_SIZE;
1166
1167 return 0;
1168}
1169
1170static int xenvif_tx_submit(struct xenvif_queue *queue)
1171{
1172 struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1173 struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1174 struct sk_buff *skb;
1175 int work_done = 0;
1176
1177 while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1178 struct xen_netif_tx_request *txp;
1179 u16 pending_idx;
1180
1181 pending_idx = copy_pending_idx(skb, 0);
1182 txp = &queue->pending_tx_info[pending_idx].req;
1183
1184 /* Check the remap error code. */
1185 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1186 /* If there was an error, xenvif_tx_check_gop is
1187 * expected to release all the frags which were mapped,
1188 * so kfree_skb shouldn't do it again
1189 */
1190 skb_shinfo(skb)->nr_frags = 0;
1191 if (skb_has_frag_list(skb)) {
1192 struct sk_buff *nskb =
1193 skb_shinfo(skb)->frag_list;
1194 skb_shinfo(nskb)->nr_frags = 0;
1195 }
1196 kfree_skb(skb);
1197 continue;
1198 }
1199
1200 if (txp->flags & XEN_NETTXF_csum_blank)
1201 skb->ip_summed = CHECKSUM_PARTIAL;
1202 else if (txp->flags & XEN_NETTXF_data_validated)
1203 skb->ip_summed = CHECKSUM_UNNECESSARY;
1204
1205 xenvif_fill_frags(queue, skb);
1206
1207 if (unlikely(skb_has_frag_list(skb))) {
1208 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1209 xenvif_skb_zerocopy_prepare(queue, nskb);
1210 if (xenvif_handle_frag_list(queue, skb)) {
1211 if (net_ratelimit())
1212 netdev_err(queue->vif->dev,
1213 "Not enough memory to consolidate frag_list!\n");
1214 xenvif_skb_zerocopy_prepare(queue, skb);
1215 kfree_skb(skb);
1216 continue;
1217 }
1218 /* Copied all the bits from the frag list -- free it. */
1219 skb_frag_list_init(skb);
1220 kfree_skb(nskb);
1221 }
1222
1223 skb->dev = queue->vif->dev;
1224 skb->protocol = eth_type_trans(skb, skb->dev);
1225 skb_reset_network_header(skb);
1226
1227 if (checksum_setup(queue, skb)) {
1228 netdev_dbg(queue->vif->dev,
1229 "Can't setup checksum in net_tx_action\n");
1230 /* We have to set this flag to trigger the callback */
1231 if (skb_shinfo(skb)->destructor_arg)
1232 xenvif_skb_zerocopy_prepare(queue, skb);
1233 kfree_skb(skb);
1234 continue;
1235 }
1236
1237 skb_probe_transport_header(skb);
1238
1239 /* If the packet is GSO then we will have just set up the
1240 * transport header offset in checksum_setup so it's now
1241 * straightforward to calculate gso_segs.
1242 */
1243 if (skb_is_gso(skb)) {
1244 int mss, hdrlen;
1245
1246 /* GSO implies having the L4 header. */
1247 WARN_ON_ONCE(!skb_transport_header_was_set(skb));
1248 if (unlikely(!skb_transport_header_was_set(skb))) {
1249 kfree_skb(skb);
1250 continue;
1251 }
1252
1253 mss = skb_shinfo(skb)->gso_size;
1254 hdrlen = skb_tcp_all_headers(skb);
1255
1256 skb_shinfo(skb)->gso_segs =
1257 DIV_ROUND_UP(skb->len - hdrlen, mss);
1258 }
1259
1260 queue->stats.rx_bytes += skb->len;
1261 queue->stats.rx_packets++;
1262
1263 work_done++;
1264
1265 /* Set this flag right before netif_receive_skb, otherwise
1266 * someone might think this packet already left netback, and
1267 * do a skb_copy_ubufs while we are still in control of the
1268 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1269 */
1270 if (skb_shinfo(skb)->destructor_arg) {
1271 xenvif_skb_zerocopy_prepare(queue, skb);
1272 queue->stats.tx_zerocopy_sent++;
1273 }
1274
1275 netif_receive_skb(skb);
1276 }
1277
1278 return work_done;
1279}
1280
1281void xenvif_zerocopy_callback(struct sk_buff *skb, struct ubuf_info *ubuf_base,
1282 bool zerocopy_success)
1283{
1284 unsigned long flags;
1285 pending_ring_idx_t index;
1286 struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
1287 struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1288
1289 /* This is the only place where we grab this lock, to protect callbacks
1290 * from each other.
1291 */
1292 spin_lock_irqsave(&queue->callback_lock, flags);
1293 do {
1294 u16 pending_idx = ubuf->desc;
1295 ubuf = (struct ubuf_info_msgzc *) ubuf->ctx;
1296 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1297 MAX_PENDING_REQS);
1298 index = pending_index(queue->dealloc_prod);
1299 queue->dealloc_ring[index] = pending_idx;
1300 /* Sync with xenvif_tx_dealloc_action:
1301 * insert idx then incr producer.
1302 */
1303 smp_wmb();
1304 queue->dealloc_prod++;
1305 } while (ubuf);
1306 spin_unlock_irqrestore(&queue->callback_lock, flags);
1307
1308 if (likely(zerocopy_success))
1309 queue->stats.tx_zerocopy_success++;
1310 else
1311 queue->stats.tx_zerocopy_fail++;
1312 xenvif_skb_zerocopy_complete(queue);
1313}
1314
1315static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1316{
1317 struct gnttab_unmap_grant_ref *gop;
1318 pending_ring_idx_t dc, dp;
1319 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1320 unsigned int i = 0;
1321
1322 dc = queue->dealloc_cons;
1323 gop = queue->tx_unmap_ops;
1324
1325 /* Free up any grants we have finished using */
1326 do {
1327 dp = queue->dealloc_prod;
1328
1329 /* Ensure we see all indices enqueued by all
1330 * xenvif_zerocopy_callback().
1331 */
1332 smp_rmb();
1333
1334 while (dc != dp) {
1335 BUG_ON(gop - queue->tx_unmap_ops >= MAX_PENDING_REQS);
1336 pending_idx =
1337 queue->dealloc_ring[pending_index(dc++)];
1338
1339 pending_idx_release[gop - queue->tx_unmap_ops] =
1340 pending_idx;
1341 queue->pages_to_unmap[gop - queue->tx_unmap_ops] =
1342 queue->mmap_pages[pending_idx];
1343 gnttab_set_unmap_op(gop,
1344 idx_to_kaddr(queue, pending_idx),
1345 GNTMAP_host_map,
1346 queue->grant_tx_handle[pending_idx]);
1347 xenvif_grant_handle_reset(queue, pending_idx);
1348 ++gop;
1349 }
1350
1351 } while (dp != queue->dealloc_prod);
1352
1353 queue->dealloc_cons = dc;
1354
1355 if (gop - queue->tx_unmap_ops > 0) {
1356 int ret;
1357 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1358 NULL,
1359 queue->pages_to_unmap,
1360 gop - queue->tx_unmap_ops);
1361 if (ret) {
1362 netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tu ret %d\n",
1363 gop - queue->tx_unmap_ops, ret);
1364 for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1365 if (gop[i].status != GNTST_okay)
1366 netdev_err(queue->vif->dev,
1367 " host_addr: 0x%llx handle: 0x%x status: %d\n",
1368 gop[i].host_addr,
1369 gop[i].handle,
1370 gop[i].status);
1371 }
1372 BUG();
1373 }
1374 }
1375
1376 for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1377 xenvif_idx_release(queue, pending_idx_release[i],
1378 XEN_NETIF_RSP_OKAY);
1379}
1380
1381
1382/* Called after netfront has transmitted */
1383int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1384{
1385 unsigned nr_mops = 0, nr_cops = 0;
1386 int work_done, ret;
1387
1388 if (unlikely(!tx_work_todo(queue)))
1389 return 0;
1390
1391 xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1392
1393 if (nr_cops == 0)
1394 return 0;
1395
1396 gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1397 if (nr_mops != 0) {
1398 ret = gnttab_map_refs(queue->tx_map_ops,
1399 NULL,
1400 queue->pages_to_map,
1401 nr_mops);
1402 if (ret) {
1403 unsigned int i;
1404
1405 netdev_err(queue->vif->dev, "Map fail: nr %u ret %d\n",
1406 nr_mops, ret);
1407 for (i = 0; i < nr_mops; ++i)
1408 WARN_ON_ONCE(queue->tx_map_ops[i].status ==
1409 GNTST_okay);
1410 }
1411 }
1412
1413 work_done = xenvif_tx_submit(queue);
1414
1415 return work_done;
1416}
1417
1418static void _make_tx_response(struct xenvif_queue *queue,
1419 const struct xen_netif_tx_request *txp,
1420 unsigned int extra_count,
1421 s8 status)
1422{
1423 RING_IDX i = queue->tx.rsp_prod_pvt;
1424 struct xen_netif_tx_response *resp;
1425
1426 resp = RING_GET_RESPONSE(&queue->tx, i);
1427 resp->id = txp->id;
1428 resp->status = status;
1429
1430 while (extra_count-- != 0)
1431 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1432
1433 queue->tx.rsp_prod_pvt = ++i;
1434}
1435
1436static void push_tx_responses(struct xenvif_queue *queue)
1437{
1438 int notify;
1439
1440 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1441 if (notify)
1442 notify_remote_via_irq(queue->tx_irq);
1443}
1444
1445static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1446 s8 status)
1447{
1448 struct pending_tx_info *pending_tx_info;
1449 pending_ring_idx_t index;
1450 unsigned long flags;
1451
1452 pending_tx_info = &queue->pending_tx_info[pending_idx];
1453
1454 spin_lock_irqsave(&queue->response_lock, flags);
1455
1456 _make_tx_response(queue, &pending_tx_info->req,
1457 pending_tx_info->extra_count, status);
1458
1459 /* Release the pending index before pusing the Tx response so
1460 * its available before a new Tx request is pushed by the
1461 * frontend.
1462 */
1463 index = pending_index(queue->pending_prod++);
1464 queue->pending_ring[index] = pending_idx;
1465
1466 push_tx_responses(queue);
1467
1468 spin_unlock_irqrestore(&queue->response_lock, flags);
1469}
1470
1471static void make_tx_response(struct xenvif_queue *queue,
1472 const struct xen_netif_tx_request *txp,
1473 unsigned int extra_count,
1474 s8 status)
1475{
1476 unsigned long flags;
1477
1478 spin_lock_irqsave(&queue->response_lock, flags);
1479
1480 _make_tx_response(queue, txp, extra_count, status);
1481 push_tx_responses(queue);
1482
1483 spin_unlock_irqrestore(&queue->response_lock, flags);
1484}
1485
1486static void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1487{
1488 int ret;
1489 struct gnttab_unmap_grant_ref tx_unmap_op;
1490
1491 gnttab_set_unmap_op(&tx_unmap_op,
1492 idx_to_kaddr(queue, pending_idx),
1493 GNTMAP_host_map,
1494 queue->grant_tx_handle[pending_idx]);
1495 xenvif_grant_handle_reset(queue, pending_idx);
1496
1497 ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1498 &queue->mmap_pages[pending_idx], 1);
1499 if (ret) {
1500 netdev_err(queue->vif->dev,
1501 "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: 0x%x status: %d\n",
1502 ret,
1503 pending_idx,
1504 tx_unmap_op.host_addr,
1505 tx_unmap_op.handle,
1506 tx_unmap_op.status);
1507 BUG();
1508 }
1509}
1510
1511static inline int tx_work_todo(struct xenvif_queue *queue)
1512{
1513 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1514 return 1;
1515
1516 return 0;
1517}
1518
1519static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1520{
1521 return queue->dealloc_cons != queue->dealloc_prod;
1522}
1523
1524void xenvif_unmap_frontend_data_rings(struct xenvif_queue *queue)
1525{
1526 if (queue->tx.sring)
1527 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1528 queue->tx.sring);
1529 if (queue->rx.sring)
1530 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1531 queue->rx.sring);
1532}
1533
1534int xenvif_map_frontend_data_rings(struct xenvif_queue *queue,
1535 grant_ref_t tx_ring_ref,
1536 grant_ref_t rx_ring_ref)
1537{
1538 void *addr;
1539 struct xen_netif_tx_sring *txs;
1540 struct xen_netif_rx_sring *rxs;
1541 RING_IDX rsp_prod, req_prod;
1542 int err;
1543
1544 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1545 &tx_ring_ref, 1, &addr);
1546 if (err)
1547 goto err;
1548
1549 txs = (struct xen_netif_tx_sring *)addr;
1550 rsp_prod = READ_ONCE(txs->rsp_prod);
1551 req_prod = READ_ONCE(txs->req_prod);
1552
1553 BACK_RING_ATTACH(&queue->tx, txs, rsp_prod, XEN_PAGE_SIZE);
1554
1555 err = -EIO;
1556 if (req_prod - rsp_prod > RING_SIZE(&queue->tx))
1557 goto err;
1558
1559 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1560 &rx_ring_ref, 1, &addr);
1561 if (err)
1562 goto err;
1563
1564 rxs = (struct xen_netif_rx_sring *)addr;
1565 rsp_prod = READ_ONCE(rxs->rsp_prod);
1566 req_prod = READ_ONCE(rxs->req_prod);
1567
1568 BACK_RING_ATTACH(&queue->rx, rxs, rsp_prod, XEN_PAGE_SIZE);
1569
1570 err = -EIO;
1571 if (req_prod - rsp_prod > RING_SIZE(&queue->rx))
1572 goto err;
1573
1574 return 0;
1575
1576err:
1577 xenvif_unmap_frontend_data_rings(queue);
1578 return err;
1579}
1580
1581static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
1582{
1583 /* Dealloc thread must remain running until all inflight
1584 * packets complete.
1585 */
1586 return kthread_should_stop() &&
1587 !atomic_read(&queue->inflight_packets);
1588}
1589
1590int xenvif_dealloc_kthread(void *data)
1591{
1592 struct xenvif_queue *queue = data;
1593
1594 for (;;) {
1595 wait_event_interruptible(queue->dealloc_wq,
1596 tx_dealloc_work_todo(queue) ||
1597 xenvif_dealloc_kthread_should_stop(queue));
1598 if (xenvif_dealloc_kthread_should_stop(queue))
1599 break;
1600
1601 xenvif_tx_dealloc_action(queue);
1602 cond_resched();
1603 }
1604
1605 /* Unmap anything remaining*/
1606 if (tx_dealloc_work_todo(queue))
1607 xenvif_tx_dealloc_action(queue);
1608
1609 return 0;
1610}
1611
1612static void make_ctrl_response(struct xenvif *vif,
1613 const struct xen_netif_ctrl_request *req,
1614 u32 status, u32 data)
1615{
1616 RING_IDX idx = vif->ctrl.rsp_prod_pvt;
1617 struct xen_netif_ctrl_response rsp = {
1618 .id = req->id,
1619 .type = req->type,
1620 .status = status,
1621 .data = data,
1622 };
1623
1624 *RING_GET_RESPONSE(&vif->ctrl, idx) = rsp;
1625 vif->ctrl.rsp_prod_pvt = ++idx;
1626}
1627
1628static void push_ctrl_response(struct xenvif *vif)
1629{
1630 int notify;
1631
1632 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->ctrl, notify);
1633 if (notify)
1634 notify_remote_via_irq(vif->ctrl_irq);
1635}
1636
1637static void process_ctrl_request(struct xenvif *vif,
1638 const struct xen_netif_ctrl_request *req)
1639{
1640 u32 status = XEN_NETIF_CTRL_STATUS_NOT_SUPPORTED;
1641 u32 data = 0;
1642
1643 switch (req->type) {
1644 case XEN_NETIF_CTRL_TYPE_SET_HASH_ALGORITHM:
1645 status = xenvif_set_hash_alg(vif, req->data[0]);
1646 break;
1647
1648 case XEN_NETIF_CTRL_TYPE_GET_HASH_FLAGS:
1649 status = xenvif_get_hash_flags(vif, &data);
1650 break;
1651
1652 case XEN_NETIF_CTRL_TYPE_SET_HASH_FLAGS:
1653 status = xenvif_set_hash_flags(vif, req->data[0]);
1654 break;
1655
1656 case XEN_NETIF_CTRL_TYPE_SET_HASH_KEY:
1657 status = xenvif_set_hash_key(vif, req->data[0],
1658 req->data[1]);
1659 break;
1660
1661 case XEN_NETIF_CTRL_TYPE_GET_HASH_MAPPING_SIZE:
1662 status = XEN_NETIF_CTRL_STATUS_SUCCESS;
1663 data = XEN_NETBK_MAX_HASH_MAPPING_SIZE;
1664 break;
1665
1666 case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING_SIZE:
1667 status = xenvif_set_hash_mapping_size(vif,
1668 req->data[0]);
1669 break;
1670
1671 case XEN_NETIF_CTRL_TYPE_SET_HASH_MAPPING:
1672 status = xenvif_set_hash_mapping(vif, req->data[0],
1673 req->data[1],
1674 req->data[2]);
1675 break;
1676
1677 default:
1678 break;
1679 }
1680
1681 make_ctrl_response(vif, req, status, data);
1682 push_ctrl_response(vif);
1683}
1684
1685static void xenvif_ctrl_action(struct xenvif *vif)
1686{
1687 for (;;) {
1688 RING_IDX req_prod, req_cons;
1689
1690 req_prod = vif->ctrl.sring->req_prod;
1691 req_cons = vif->ctrl.req_cons;
1692
1693 /* Make sure we can see requests before we process them. */
1694 rmb();
1695
1696 if (req_cons == req_prod)
1697 break;
1698
1699 while (req_cons != req_prod) {
1700 struct xen_netif_ctrl_request req;
1701
1702 RING_COPY_REQUEST(&vif->ctrl, req_cons, &req);
1703 req_cons++;
1704
1705 process_ctrl_request(vif, &req);
1706 }
1707
1708 vif->ctrl.req_cons = req_cons;
1709 vif->ctrl.sring->req_event = req_cons + 1;
1710 }
1711}
1712
1713static bool xenvif_ctrl_work_todo(struct xenvif *vif)
1714{
1715 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->ctrl)))
1716 return true;
1717
1718 return false;
1719}
1720
1721irqreturn_t xenvif_ctrl_irq_fn(int irq, void *data)
1722{
1723 struct xenvif *vif = data;
1724 unsigned int eoi_flag = XEN_EOI_FLAG_SPURIOUS;
1725
1726 while (xenvif_ctrl_work_todo(vif)) {
1727 xenvif_ctrl_action(vif);
1728 eoi_flag = 0;
1729 }
1730
1731 xen_irq_lateeoi(irq, eoi_flag);
1732
1733 return IRQ_HANDLED;
1734}
1735
1736static int __init netback_init(void)
1737{
1738 int rc = 0;
1739
1740 if (!xen_domain())
1741 return -ENODEV;
1742
1743 /* Allow as many queues as there are CPUs but max. 8 if user has not
1744 * specified a value.
1745 */
1746 if (xenvif_max_queues == 0)
1747 xenvif_max_queues = min_t(unsigned int, MAX_QUEUES_DEFAULT,
1748 num_online_cpus());
1749
1750 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1751 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1752 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1753 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1754 }
1755
1756 rc = xenvif_xenbus_init();
1757 if (rc)
1758 goto failed_init;
1759
1760#ifdef CONFIG_DEBUG_FS
1761 xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
1762#endif /* CONFIG_DEBUG_FS */
1763
1764 return 0;
1765
1766failed_init:
1767 return rc;
1768}
1769
1770module_init(netback_init);
1771
1772static void __exit netback_fini(void)
1773{
1774#ifdef CONFIG_DEBUG_FS
1775 debugfs_remove_recursive(xen_netback_dbg_root);
1776#endif /* CONFIG_DEBUG_FS */
1777 xenvif_xenbus_fini();
1778}
1779module_exit(netback_fini);
1780
1781MODULE_DESCRIPTION("Xen backend network device module");
1782MODULE_LICENSE("Dual BSD/GPL");
1783MODULE_ALIAS("xen-backend:vif");
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 */
55bool separate_tx_rx_irq = 1;
56module_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 stuck there
60 */
61unsigned int rx_drain_timeout_msecs = 10000;
62module_param(rx_drain_timeout_msecs, uint, 0444);
63unsigned 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
70static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
71module_param(fatal_skb_slots, uint, 0444);
72
73static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
74 u8 status);
75
76static void make_tx_response(struct xenvif *vif,
77 struct xen_netif_tx_request *txp,
78 s8 st);
79
80static inline int tx_work_todo(struct xenvif *vif);
81static inline int rx_work_todo(struct xenvif *vif);
82
83static 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
90static inline unsigned long idx_to_pfn(struct xenvif *vif,
91 u16 idx)
92{
93 return page_to_pfn(vif->mmap_pages[idx]);
94}
95
96static 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#define callback_param(vif, pending_idx) \
103 (vif->pending_tx_info[pending_idx].callback_struct)
104
105/* Find the containing VIF's structure from a pointer in pending_tx_info array
106 */
107static inline struct xenvif *ubuf_to_vif(const struct ubuf_info *ubuf)
108{
109 u16 pending_idx = ubuf->desc;
110 struct pending_tx_info *temp =
111 container_of(ubuf, struct pending_tx_info, callback_struct);
112 return container_of(temp - pending_idx,
113 struct xenvif,
114 pending_tx_info[0]);
115}
116
117/* This is a miniumum size for the linear area to avoid lots of
118 * calls to __pskb_pull_tail() as we set up checksum offsets. The
119 * value 128 was chosen as it covers all IPv4 and most likely
120 * IPv6 headers.
121 */
122#define PKT_PROT_LEN 128
123
124static u16 frag_get_pending_idx(skb_frag_t *frag)
125{
126 return (u16)frag->page_offset;
127}
128
129static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
130{
131 frag->page_offset = pending_idx;
132}
133
134static inline pending_ring_idx_t pending_index(unsigned i)
135{
136 return i & (MAX_PENDING_REQS-1);
137}
138
139bool xenvif_rx_ring_slots_available(struct xenvif *vif, int needed)
140{
141 RING_IDX prod, cons;
142
143 do {
144 prod = vif->rx.sring->req_prod;
145 cons = vif->rx.req_cons;
146
147 if (prod - cons >= needed)
148 return true;
149
150 vif->rx.sring->req_event = prod + 1;
151
152 /* Make sure event is visible before we check prod
153 * again.
154 */
155 mb();
156 } while (vif->rx.sring->req_prod != prod);
157
158 return false;
159}
160
161/*
162 * Returns true if we should start a new receive buffer instead of
163 * adding 'size' bytes to a buffer which currently contains 'offset'
164 * bytes.
165 */
166static bool start_new_rx_buffer(int offset, unsigned long size, int head)
167{
168 /* simple case: we have completely filled the current buffer. */
169 if (offset == MAX_BUFFER_OFFSET)
170 return true;
171
172 /*
173 * complex case: start a fresh buffer if the current frag
174 * would overflow the current buffer but only if:
175 * (i) this frag would fit completely in the next buffer
176 * and (ii) there is already some data in the current buffer
177 * and (iii) this is not the head buffer.
178 *
179 * Where:
180 * - (i) stops us splitting a frag into two copies
181 * unless the frag is too large for a single buffer.
182 * - (ii) stops us from leaving a buffer pointlessly empty.
183 * - (iii) stops us leaving the first buffer
184 * empty. Strictly speaking this is already covered
185 * by (ii) but is explicitly checked because
186 * netfront relies on the first buffer being
187 * non-empty and can crash otherwise.
188 *
189 * This means we will effectively linearise small
190 * frags but do not needlessly split large buffers
191 * into multiple copies tend to give large frags their
192 * own buffers as before.
193 */
194 BUG_ON(size > MAX_BUFFER_OFFSET);
195 if ((offset + size > MAX_BUFFER_OFFSET) && offset && !head)
196 return true;
197
198 return false;
199}
200
201struct netrx_pending_operations {
202 unsigned copy_prod, copy_cons;
203 unsigned meta_prod, meta_cons;
204 struct gnttab_copy *copy;
205 struct xenvif_rx_meta *meta;
206 int copy_off;
207 grant_ref_t copy_gref;
208};
209
210static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif *vif,
211 struct netrx_pending_operations *npo)
212{
213 struct xenvif_rx_meta *meta;
214 struct xen_netif_rx_request *req;
215
216 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
217
218 meta = npo->meta + npo->meta_prod++;
219 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
220 meta->gso_size = 0;
221 meta->size = 0;
222 meta->id = req->id;
223
224 npo->copy_off = 0;
225 npo->copy_gref = req->gref;
226
227 return meta;
228}
229
230/*
231 * Set up the grant operations for this fragment. If it's a flipping
232 * interface, we also set up the unmap request from here.
233 */
234static void xenvif_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
235 struct netrx_pending_operations *npo,
236 struct page *page, unsigned long size,
237 unsigned long offset, int *head,
238 struct xenvif *foreign_vif,
239 grant_ref_t foreign_gref)
240{
241 struct gnttab_copy *copy_gop;
242 struct xenvif_rx_meta *meta;
243 unsigned long bytes;
244 int gso_type = XEN_NETIF_GSO_TYPE_NONE;
245
246 /* Data must not cross a page boundary. */
247 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
248
249 meta = npo->meta + npo->meta_prod - 1;
250
251 /* Skip unused frames from start of page */
252 page += offset >> PAGE_SHIFT;
253 offset &= ~PAGE_MASK;
254
255 while (size > 0) {
256 BUG_ON(offset >= PAGE_SIZE);
257 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
258
259 bytes = PAGE_SIZE - offset;
260
261 if (bytes > size)
262 bytes = size;
263
264 if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
265 /*
266 * Netfront requires there to be some data in the head
267 * buffer.
268 */
269 BUG_ON(*head);
270
271 meta = get_next_rx_buffer(vif, npo);
272 }
273
274 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
275 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
276
277 copy_gop = npo->copy + npo->copy_prod++;
278 copy_gop->flags = GNTCOPY_dest_gref;
279 copy_gop->len = bytes;
280
281 if (foreign_vif) {
282 copy_gop->source.domid = foreign_vif->domid;
283 copy_gop->source.u.ref = foreign_gref;
284 copy_gop->flags |= GNTCOPY_source_gref;
285 } else {
286 copy_gop->source.domid = DOMID_SELF;
287 copy_gop->source.u.gmfn =
288 virt_to_mfn(page_address(page));
289 }
290 copy_gop->source.offset = offset;
291
292 copy_gop->dest.domid = vif->domid;
293 copy_gop->dest.offset = npo->copy_off;
294 copy_gop->dest.u.ref = npo->copy_gref;
295
296 npo->copy_off += bytes;
297 meta->size += bytes;
298
299 offset += bytes;
300 size -= bytes;
301
302 /* Next frame */
303 if (offset == PAGE_SIZE && size) {
304 BUG_ON(!PageCompound(page));
305 page++;
306 offset = 0;
307 }
308
309 /* Leave a gap for the GSO descriptor. */
310 if (skb_is_gso(skb)) {
311 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
312 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
313 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
314 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
315 }
316
317 if (*head && ((1 << gso_type) & vif->gso_mask))
318 vif->rx.req_cons++;
319
320 *head = 0; /* There must be something in this buffer now. */
321
322 }
323}
324
325/*
326 * Find the grant ref for a given frag in a chain of struct ubuf_info's
327 * skb: the skb itself
328 * i: the frag's number
329 * ubuf: a pointer to an element in the chain. It should not be NULL
330 *
331 * Returns a pointer to the element in the chain where the page were found. If
332 * not found, returns NULL.
333 * See the definition of callback_struct in common.h for more details about
334 * the chain.
335 */
336static const struct ubuf_info *xenvif_find_gref(const struct sk_buff *const skb,
337 const int i,
338 const struct ubuf_info *ubuf)
339{
340 struct xenvif *foreign_vif = ubuf_to_vif(ubuf);
341
342 do {
343 u16 pending_idx = ubuf->desc;
344
345 if (skb_shinfo(skb)->frags[i].page.p ==
346 foreign_vif->mmap_pages[pending_idx])
347 break;
348 ubuf = (struct ubuf_info *) ubuf->ctx;
349 } while (ubuf);
350
351 return ubuf;
352}
353
354/*
355 * Prepare an SKB to be transmitted to the frontend.
356 *
357 * This function is responsible for allocating grant operations, meta
358 * structures, etc.
359 *
360 * It returns the number of meta structures consumed. The number of
361 * ring slots used is always equal to the number of meta slots used
362 * plus the number of GSO descriptors used. Currently, we use either
363 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
364 * frontend-side LRO).
365 */
366static int xenvif_gop_skb(struct sk_buff *skb,
367 struct netrx_pending_operations *npo)
368{
369 struct xenvif *vif = netdev_priv(skb->dev);
370 int nr_frags = skb_shinfo(skb)->nr_frags;
371 int i;
372 struct xen_netif_rx_request *req;
373 struct xenvif_rx_meta *meta;
374 unsigned char *data;
375 int head = 1;
376 int old_meta_prod;
377 int gso_type;
378 const struct ubuf_info *ubuf = skb_shinfo(skb)->destructor_arg;
379 const struct ubuf_info *const head_ubuf = ubuf;
380
381 old_meta_prod = npo->meta_prod;
382
383 gso_type = XEN_NETIF_GSO_TYPE_NONE;
384 if (skb_is_gso(skb)) {
385 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
386 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
387 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
388 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
389 }
390
391 /* Set up a GSO prefix descriptor, if necessary */
392 if ((1 << gso_type) & vif->gso_prefix_mask) {
393 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
394 meta = npo->meta + npo->meta_prod++;
395 meta->gso_type = gso_type;
396 meta->gso_size = skb_shinfo(skb)->gso_size;
397 meta->size = 0;
398 meta->id = req->id;
399 }
400
401 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
402 meta = npo->meta + npo->meta_prod++;
403
404 if ((1 << gso_type) & vif->gso_mask) {
405 meta->gso_type = gso_type;
406 meta->gso_size = skb_shinfo(skb)->gso_size;
407 } else {
408 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
409 meta->gso_size = 0;
410 }
411
412 meta->size = 0;
413 meta->id = req->id;
414 npo->copy_off = 0;
415 npo->copy_gref = req->gref;
416
417 data = skb->data;
418 while (data < skb_tail_pointer(skb)) {
419 unsigned int offset = offset_in_page(data);
420 unsigned int len = PAGE_SIZE - offset;
421
422 if (data + len > skb_tail_pointer(skb))
423 len = skb_tail_pointer(skb) - data;
424
425 xenvif_gop_frag_copy(vif, skb, npo,
426 virt_to_page(data), len, offset, &head,
427 NULL,
428 0);
429 data += len;
430 }
431
432 for (i = 0; i < nr_frags; i++) {
433 /* This variable also signals whether foreign_gref has a real
434 * value or not.
435 */
436 struct xenvif *foreign_vif = NULL;
437 grant_ref_t foreign_gref;
438
439 if ((skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY) &&
440 (ubuf->callback == &xenvif_zerocopy_callback)) {
441 const struct ubuf_info *const startpoint = ubuf;
442
443 /* Ideally ubuf points to the chain element which
444 * belongs to this frag. Or if frags were removed from
445 * the beginning, then shortly before it.
446 */
447 ubuf = xenvif_find_gref(skb, i, ubuf);
448
449 /* Try again from the beginning of the list, if we
450 * haven't tried from there. This only makes sense in
451 * the unlikely event of reordering the original frags.
452 * For injected local pages it's an unnecessary second
453 * run.
454 */
455 if (unlikely(!ubuf) && startpoint != head_ubuf)
456 ubuf = xenvif_find_gref(skb, i, head_ubuf);
457
458 if (likely(ubuf)) {
459 u16 pending_idx = ubuf->desc;
460
461 foreign_vif = ubuf_to_vif(ubuf);
462 foreign_gref = foreign_vif->pending_tx_info[pending_idx].req.gref;
463 /* Just a safety measure. If this was the last
464 * element on the list, the for loop will
465 * iterate again if a local page were added to
466 * the end. Using head_ubuf here prevents the
467 * second search on the chain. Or the original
468 * frags changed order, but that's less likely.
469 * In any way, ubuf shouldn't be NULL.
470 */
471 ubuf = ubuf->ctx ?
472 (struct ubuf_info *) ubuf->ctx :
473 head_ubuf;
474 } else
475 /* This frag was a local page, added to the
476 * array after the skb left netback.
477 */
478 ubuf = head_ubuf;
479 }
480 xenvif_gop_frag_copy(vif, skb, npo,
481 skb_frag_page(&skb_shinfo(skb)->frags[i]),
482 skb_frag_size(&skb_shinfo(skb)->frags[i]),
483 skb_shinfo(skb)->frags[i].page_offset,
484 &head,
485 foreign_vif,
486 foreign_vif ? foreign_gref : UINT_MAX);
487 }
488
489 return npo->meta_prod - old_meta_prod;
490}
491
492/*
493 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
494 * used to set up the operations on the top of
495 * netrx_pending_operations, which have since been done. Check that
496 * they didn't give any errors and advance over them.
497 */
498static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
499 struct netrx_pending_operations *npo)
500{
501 struct gnttab_copy *copy_op;
502 int status = XEN_NETIF_RSP_OKAY;
503 int i;
504
505 for (i = 0; i < nr_meta_slots; i++) {
506 copy_op = npo->copy + npo->copy_cons++;
507 if (copy_op->status != GNTST_okay) {
508 netdev_dbg(vif->dev,
509 "Bad status %d from copy to DOM%d.\n",
510 copy_op->status, vif->domid);
511 status = XEN_NETIF_RSP_ERROR;
512 }
513 }
514
515 return status;
516}
517
518static void xenvif_add_frag_responses(struct xenvif *vif, int status,
519 struct xenvif_rx_meta *meta,
520 int nr_meta_slots)
521{
522 int i;
523 unsigned long offset;
524
525 /* No fragments used */
526 if (nr_meta_slots <= 1)
527 return;
528
529 nr_meta_slots--;
530
531 for (i = 0; i < nr_meta_slots; i++) {
532 int flags;
533 if (i == nr_meta_slots - 1)
534 flags = 0;
535 else
536 flags = XEN_NETRXF_more_data;
537
538 offset = 0;
539 make_rx_response(vif, meta[i].id, status, offset,
540 meta[i].size, flags);
541 }
542}
543
544struct xenvif_rx_cb {
545 int meta_slots_used;
546};
547
548#define XENVIF_RX_CB(skb) ((struct xenvif_rx_cb *)(skb)->cb)
549
550void xenvif_kick_thread(struct xenvif *vif)
551{
552 wake_up(&vif->wq);
553}
554
555static void xenvif_rx_action(struct xenvif *vif)
556{
557 s8 status;
558 u16 flags;
559 struct xen_netif_rx_response *resp;
560 struct sk_buff_head rxq;
561 struct sk_buff *skb;
562 LIST_HEAD(notify);
563 int ret;
564 unsigned long offset;
565 bool need_to_notify = false;
566
567 struct netrx_pending_operations npo = {
568 .copy = vif->grant_copy_op,
569 .meta = vif->meta,
570 };
571
572 skb_queue_head_init(&rxq);
573
574 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL) {
575 RING_IDX max_slots_needed;
576 RING_IDX old_req_cons;
577 RING_IDX ring_slots_used;
578 int i;
579
580 /* We need a cheap worse case estimate for the number of
581 * slots we'll use.
582 */
583
584 max_slots_needed = DIV_ROUND_UP(offset_in_page(skb->data) +
585 skb_headlen(skb),
586 PAGE_SIZE);
587 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
588 unsigned int size;
589 unsigned int offset;
590
591 size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
592 offset = skb_shinfo(skb)->frags[i].page_offset;
593
594 /* For a worse-case estimate we need to factor in
595 * the fragment page offset as this will affect the
596 * number of times xenvif_gop_frag_copy() will
597 * call start_new_rx_buffer().
598 */
599 max_slots_needed += DIV_ROUND_UP(offset + size,
600 PAGE_SIZE);
601 }
602
603 /* To avoid the estimate becoming too pessimal for some
604 * frontends that limit posted rx requests, cap the estimate
605 * at MAX_SKB_FRAGS.
606 */
607 if (max_slots_needed > MAX_SKB_FRAGS)
608 max_slots_needed = MAX_SKB_FRAGS;
609
610 /* We may need one more slot for GSO metadata */
611 if (skb_is_gso(skb) &&
612 (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4 ||
613 skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6))
614 max_slots_needed++;
615
616 /* If the skb may not fit then bail out now */
617 if (!xenvif_rx_ring_slots_available(vif, max_slots_needed)) {
618 skb_queue_head(&vif->rx_queue, skb);
619 need_to_notify = true;
620 vif->rx_last_skb_slots = max_slots_needed;
621 break;
622 } else
623 vif->rx_last_skb_slots = 0;
624
625 old_req_cons = vif->rx.req_cons;
626 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo);
627 ring_slots_used = vif->rx.req_cons - old_req_cons;
628
629 BUG_ON(ring_slots_used > max_slots_needed);
630
631 __skb_queue_tail(&rxq, skb);
632 }
633
634 BUG_ON(npo.meta_prod > ARRAY_SIZE(vif->meta));
635
636 if (!npo.copy_prod)
637 goto done;
638
639 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
640 gnttab_batch_copy(vif->grant_copy_op, npo.copy_prod);
641
642 while ((skb = __skb_dequeue(&rxq)) != NULL) {
643
644 if ((1 << vif->meta[npo.meta_cons].gso_type) &
645 vif->gso_prefix_mask) {
646 resp = RING_GET_RESPONSE(&vif->rx,
647 vif->rx.rsp_prod_pvt++);
648
649 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
650
651 resp->offset = vif->meta[npo.meta_cons].gso_size;
652 resp->id = vif->meta[npo.meta_cons].id;
653 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
654
655 npo.meta_cons++;
656 XENVIF_RX_CB(skb)->meta_slots_used--;
657 }
658
659
660 vif->dev->stats.tx_bytes += skb->len;
661 vif->dev->stats.tx_packets++;
662
663 status = xenvif_check_gop(vif,
664 XENVIF_RX_CB(skb)->meta_slots_used,
665 &npo);
666
667 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
668 flags = 0;
669 else
670 flags = XEN_NETRXF_more_data;
671
672 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
673 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
674 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
675 /* remote but checksummed. */
676 flags |= XEN_NETRXF_data_validated;
677
678 offset = 0;
679 resp = make_rx_response(vif, vif->meta[npo.meta_cons].id,
680 status, offset,
681 vif->meta[npo.meta_cons].size,
682 flags);
683
684 if ((1 << vif->meta[npo.meta_cons].gso_type) &
685 vif->gso_mask) {
686 struct xen_netif_extra_info *gso =
687 (struct xen_netif_extra_info *)
688 RING_GET_RESPONSE(&vif->rx,
689 vif->rx.rsp_prod_pvt++);
690
691 resp->flags |= XEN_NETRXF_extra_info;
692
693 gso->u.gso.type = vif->meta[npo.meta_cons].gso_type;
694 gso->u.gso.size = vif->meta[npo.meta_cons].gso_size;
695 gso->u.gso.pad = 0;
696 gso->u.gso.features = 0;
697
698 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
699 gso->flags = 0;
700 }
701
702 xenvif_add_frag_responses(vif, status,
703 vif->meta + npo.meta_cons + 1,
704 XENVIF_RX_CB(skb)->meta_slots_used);
705
706 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
707
708 need_to_notify |= !!ret;
709
710 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
711 dev_kfree_skb(skb);
712 }
713
714done:
715 if (need_to_notify)
716 notify_remote_via_irq(vif->rx_irq);
717}
718
719void xenvif_napi_schedule_or_enable_events(struct xenvif *vif)
720{
721 int more_to_do;
722
723 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
724
725 if (more_to_do)
726 napi_schedule(&vif->napi);
727}
728
729static void tx_add_credit(struct xenvif *vif)
730{
731 unsigned long max_burst, max_credit;
732
733 /*
734 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
735 * Otherwise the interface can seize up due to insufficient credit.
736 */
737 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
738 max_burst = min(max_burst, 131072UL);
739 max_burst = max(max_burst, vif->credit_bytes);
740
741 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
742 max_credit = vif->remaining_credit + vif->credit_bytes;
743 if (max_credit < vif->remaining_credit)
744 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
745
746 vif->remaining_credit = min(max_credit, max_burst);
747}
748
749static void tx_credit_callback(unsigned long data)
750{
751 struct xenvif *vif = (struct xenvif *)data;
752 tx_add_credit(vif);
753 xenvif_napi_schedule_or_enable_events(vif);
754}
755
756static void xenvif_tx_err(struct xenvif *vif,
757 struct xen_netif_tx_request *txp, RING_IDX end)
758{
759 RING_IDX cons = vif->tx.req_cons;
760 unsigned long flags;
761
762 do {
763 spin_lock_irqsave(&vif->response_lock, flags);
764 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
765 spin_unlock_irqrestore(&vif->response_lock, flags);
766 if (cons == end)
767 break;
768 txp = RING_GET_REQUEST(&vif->tx, cons++);
769 } while (1);
770 vif->tx.req_cons = cons;
771}
772
773static void xenvif_fatal_tx_err(struct xenvif *vif)
774{
775 netdev_err(vif->dev, "fatal error; disabling device\n");
776 vif->disabled = true;
777 xenvif_kick_thread(vif);
778}
779
780static int xenvif_count_requests(struct xenvif *vif,
781 struct xen_netif_tx_request *first,
782 struct xen_netif_tx_request *txp,
783 int work_to_do)
784{
785 RING_IDX cons = vif->tx.req_cons;
786 int slots = 0;
787 int drop_err = 0;
788 int more_data;
789
790 if (!(first->flags & XEN_NETTXF_more_data))
791 return 0;
792
793 do {
794 struct xen_netif_tx_request dropped_tx = { 0 };
795
796 if (slots >= work_to_do) {
797 netdev_err(vif->dev,
798 "Asked for %d slots but exceeds this limit\n",
799 work_to_do);
800 xenvif_fatal_tx_err(vif);
801 return -ENODATA;
802 }
803
804 /* This guest is really using too many slots and
805 * considered malicious.
806 */
807 if (unlikely(slots >= fatal_skb_slots)) {
808 netdev_err(vif->dev,
809 "Malicious frontend using %d slots, threshold %u\n",
810 slots, fatal_skb_slots);
811 xenvif_fatal_tx_err(vif);
812 return -E2BIG;
813 }
814
815 /* Xen network protocol had implicit dependency on
816 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
817 * the historical MAX_SKB_FRAGS value 18 to honor the
818 * same behavior as before. Any packet using more than
819 * 18 slots but less than fatal_skb_slots slots is
820 * dropped
821 */
822 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
823 if (net_ratelimit())
824 netdev_dbg(vif->dev,
825 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
826 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
827 drop_err = -E2BIG;
828 }
829
830 if (drop_err)
831 txp = &dropped_tx;
832
833 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
834 sizeof(*txp));
835
836 /* If the guest submitted a frame >= 64 KiB then
837 * first->size overflowed and following slots will
838 * appear to be larger than the frame.
839 *
840 * This cannot be fatal error as there are buggy
841 * frontends that do this.
842 *
843 * Consume all slots and drop the packet.
844 */
845 if (!drop_err && txp->size > first->size) {
846 if (net_ratelimit())
847 netdev_dbg(vif->dev,
848 "Invalid tx request, slot size %u > remaining size %u\n",
849 txp->size, first->size);
850 drop_err = -EIO;
851 }
852
853 first->size -= txp->size;
854 slots++;
855
856 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
857 netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
858 txp->offset, txp->size);
859 xenvif_fatal_tx_err(vif);
860 return -EINVAL;
861 }
862
863 more_data = txp->flags & XEN_NETTXF_more_data;
864
865 if (!drop_err)
866 txp++;
867
868 } while (more_data);
869
870 if (drop_err) {
871 xenvif_tx_err(vif, first, cons + slots);
872 return drop_err;
873 }
874
875 return slots;
876}
877
878
879struct xenvif_tx_cb {
880 u16 pending_idx;
881};
882
883#define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
884
885static inline void xenvif_tx_create_map_op(struct xenvif *vif,
886 u16 pending_idx,
887 struct xen_netif_tx_request *txp,
888 struct gnttab_map_grant_ref *mop)
889{
890 vif->pages_to_map[mop-vif->tx_map_ops] = vif->mmap_pages[pending_idx];
891 gnttab_set_map_op(mop, idx_to_kaddr(vif, pending_idx),
892 GNTMAP_host_map | GNTMAP_readonly,
893 txp->gref, vif->domid);
894
895 memcpy(&vif->pending_tx_info[pending_idx].req, txp,
896 sizeof(*txp));
897}
898
899static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
900{
901 struct sk_buff *skb =
902 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
903 GFP_ATOMIC | __GFP_NOWARN);
904 if (unlikely(skb == NULL))
905 return NULL;
906
907 /* Packets passed to netif_rx() must have some headroom. */
908 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
909
910 /* Initialize it here to avoid later surprises */
911 skb_shinfo(skb)->destructor_arg = NULL;
912
913 return skb;
914}
915
916static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif *vif,
917 struct sk_buff *skb,
918 struct xen_netif_tx_request *txp,
919 struct gnttab_map_grant_ref *gop)
920{
921 struct skb_shared_info *shinfo = skb_shinfo(skb);
922 skb_frag_t *frags = shinfo->frags;
923 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
924 int start;
925 pending_ring_idx_t index;
926 unsigned int nr_slots, frag_overflow = 0;
927
928 /* At this point shinfo->nr_frags is in fact the number of
929 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
930 */
931 if (shinfo->nr_frags > MAX_SKB_FRAGS) {
932 frag_overflow = shinfo->nr_frags - MAX_SKB_FRAGS;
933 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
934 shinfo->nr_frags = MAX_SKB_FRAGS;
935 }
936 nr_slots = shinfo->nr_frags;
937
938 /* Skip first skb fragment if it is on same page as header fragment. */
939 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
940
941 for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
942 shinfo->nr_frags++, txp++, gop++) {
943 index = pending_index(vif->pending_cons++);
944 pending_idx = vif->pending_ring[index];
945 xenvif_tx_create_map_op(vif, pending_idx, txp, gop);
946 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
947 }
948
949 if (frag_overflow) {
950 struct sk_buff *nskb = xenvif_alloc_skb(0);
951 if (unlikely(nskb == NULL)) {
952 if (net_ratelimit())
953 netdev_err(vif->dev,
954 "Can't allocate the frag_list skb.\n");
955 return NULL;
956 }
957
958 shinfo = skb_shinfo(nskb);
959 frags = shinfo->frags;
960
961 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
962 shinfo->nr_frags++, txp++, gop++) {
963 index = pending_index(vif->pending_cons++);
964 pending_idx = vif->pending_ring[index];
965 xenvif_tx_create_map_op(vif, pending_idx, txp, gop);
966 frag_set_pending_idx(&frags[shinfo->nr_frags],
967 pending_idx);
968 }
969
970 skb_shinfo(skb)->frag_list = nskb;
971 }
972
973 return gop;
974}
975
976static inline void xenvif_grant_handle_set(struct xenvif *vif,
977 u16 pending_idx,
978 grant_handle_t handle)
979{
980 if (unlikely(vif->grant_tx_handle[pending_idx] !=
981 NETBACK_INVALID_HANDLE)) {
982 netdev_err(vif->dev,
983 "Trying to overwrite active handle! pending_idx: %x\n",
984 pending_idx);
985 BUG();
986 }
987 vif->grant_tx_handle[pending_idx] = handle;
988}
989
990static inline void xenvif_grant_handle_reset(struct xenvif *vif,
991 u16 pending_idx)
992{
993 if (unlikely(vif->grant_tx_handle[pending_idx] ==
994 NETBACK_INVALID_HANDLE)) {
995 netdev_err(vif->dev,
996 "Trying to unmap invalid handle! pending_idx: %x\n",
997 pending_idx);
998 BUG();
999 }
1000 vif->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
1001}
1002
1003static int xenvif_tx_check_gop(struct xenvif *vif,
1004 struct sk_buff *skb,
1005 struct gnttab_map_grant_ref **gopp_map,
1006 struct gnttab_copy **gopp_copy)
1007{
1008 struct gnttab_map_grant_ref *gop_map = *gopp_map;
1009 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1010 struct skb_shared_info *shinfo = skb_shinfo(skb);
1011 int nr_frags = shinfo->nr_frags;
1012 int i, err;
1013 struct sk_buff *first_skb = NULL;
1014
1015 /* Check status of header. */
1016 err = (*gopp_copy)->status;
1017 (*gopp_copy)++;
1018 if (unlikely(err)) {
1019 if (net_ratelimit())
1020 netdev_dbg(vif->dev,
1021 "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
1022 (*gopp_copy)->status,
1023 pending_idx,
1024 (*gopp_copy)->source.u.ref);
1025 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
1026 }
1027
1028check_frags:
1029 for (i = 0; i < nr_frags; i++, gop_map++) {
1030 int j, newerr;
1031
1032 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
1033
1034 /* Check error status: if okay then remember grant handle. */
1035 newerr = gop_map->status;
1036
1037 if (likely(!newerr)) {
1038 xenvif_grant_handle_set(vif,
1039 pending_idx,
1040 gop_map->handle);
1041 /* Had a previous error? Invalidate this fragment. */
1042 if (unlikely(err))
1043 xenvif_idx_unmap(vif, pending_idx);
1044 continue;
1045 }
1046
1047 /* Error on this fragment: respond to client with an error. */
1048 if (net_ratelimit())
1049 netdev_dbg(vif->dev,
1050 "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
1051 i,
1052 gop_map->status,
1053 pending_idx,
1054 gop_map->ref);
1055 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
1056
1057 /* Not the first error? Preceding frags already invalidated. */
1058 if (err)
1059 continue;
1060 /* First error: invalidate preceding fragments. */
1061 for (j = 0; j < i; j++) {
1062 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1063 xenvif_idx_unmap(vif, pending_idx);
1064 }
1065
1066 /* Remember the error: invalidate all subsequent fragments. */
1067 err = newerr;
1068 }
1069
1070 if (skb_has_frag_list(skb)) {
1071 first_skb = skb;
1072 skb = shinfo->frag_list;
1073 shinfo = skb_shinfo(skb);
1074 nr_frags = shinfo->nr_frags;
1075
1076 goto check_frags;
1077 }
1078
1079 /* There was a mapping error in the frag_list skb. We have to unmap
1080 * the first skb's frags
1081 */
1082 if (first_skb && err) {
1083 int j;
1084 shinfo = skb_shinfo(first_skb);
1085 for (j = 0; j < shinfo->nr_frags; j++) {
1086 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1087 xenvif_idx_unmap(vif, pending_idx);
1088 }
1089 }
1090
1091 *gopp_map = gop_map;
1092 return err;
1093}
1094
1095static void xenvif_fill_frags(struct xenvif *vif, struct sk_buff *skb)
1096{
1097 struct skb_shared_info *shinfo = skb_shinfo(skb);
1098 int nr_frags = shinfo->nr_frags;
1099 int i;
1100 u16 prev_pending_idx = INVALID_PENDING_IDX;
1101
1102 for (i = 0; i < nr_frags; i++) {
1103 skb_frag_t *frag = shinfo->frags + i;
1104 struct xen_netif_tx_request *txp;
1105 struct page *page;
1106 u16 pending_idx;
1107
1108 pending_idx = frag_get_pending_idx(frag);
1109
1110 /* If this is not the first frag, chain it to the previous*/
1111 if (prev_pending_idx == INVALID_PENDING_IDX)
1112 skb_shinfo(skb)->destructor_arg =
1113 &callback_param(vif, pending_idx);
1114 else
1115 callback_param(vif, prev_pending_idx).ctx =
1116 &callback_param(vif, pending_idx);
1117
1118 callback_param(vif, pending_idx).ctx = NULL;
1119 prev_pending_idx = pending_idx;
1120
1121 txp = &vif->pending_tx_info[pending_idx].req;
1122 page = virt_to_page(idx_to_kaddr(vif, pending_idx));
1123 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1124 skb->len += txp->size;
1125 skb->data_len += txp->size;
1126 skb->truesize += txp->size;
1127
1128 /* Take an extra reference to offset network stack's put_page */
1129 get_page(vif->mmap_pages[pending_idx]);
1130 }
1131 /* FIXME: __skb_fill_page_desc set this to true because page->pfmemalloc
1132 * overlaps with "index", and "mapping" is not set. I think mapping
1133 * should be set. If delivered to local stack, it would drop this
1134 * skb in sk_filter unless the socket has the right to use it.
1135 */
1136 skb->pfmemalloc = false;
1137}
1138
1139static int xenvif_get_extras(struct xenvif *vif,
1140 struct xen_netif_extra_info *extras,
1141 int work_to_do)
1142{
1143 struct xen_netif_extra_info extra;
1144 RING_IDX cons = vif->tx.req_cons;
1145
1146 do {
1147 if (unlikely(work_to_do-- <= 0)) {
1148 netdev_err(vif->dev, "Missing extra info\n");
1149 xenvif_fatal_tx_err(vif);
1150 return -EBADR;
1151 }
1152
1153 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1154 sizeof(extra));
1155 if (unlikely(!extra.type ||
1156 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1157 vif->tx.req_cons = ++cons;
1158 netdev_err(vif->dev,
1159 "Invalid extra type: %d\n", extra.type);
1160 xenvif_fatal_tx_err(vif);
1161 return -EINVAL;
1162 }
1163
1164 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1165 vif->tx.req_cons = ++cons;
1166 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1167
1168 return work_to_do;
1169}
1170
1171static int xenvif_set_skb_gso(struct xenvif *vif,
1172 struct sk_buff *skb,
1173 struct xen_netif_extra_info *gso)
1174{
1175 if (!gso->u.gso.size) {
1176 netdev_err(vif->dev, "GSO size must not be zero.\n");
1177 xenvif_fatal_tx_err(vif);
1178 return -EINVAL;
1179 }
1180
1181 switch (gso->u.gso.type) {
1182 case XEN_NETIF_GSO_TYPE_TCPV4:
1183 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1184 break;
1185 case XEN_NETIF_GSO_TYPE_TCPV6:
1186 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1187 break;
1188 default:
1189 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1190 xenvif_fatal_tx_err(vif);
1191 return -EINVAL;
1192 }
1193
1194 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1195 /* gso_segs will be calculated later */
1196
1197 return 0;
1198}
1199
1200static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1201{
1202 bool recalculate_partial_csum = false;
1203
1204 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1205 * peers can fail to set NETRXF_csum_blank when sending a GSO
1206 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1207 * recalculate the partial checksum.
1208 */
1209 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1210 vif->rx_gso_checksum_fixup++;
1211 skb->ip_summed = CHECKSUM_PARTIAL;
1212 recalculate_partial_csum = true;
1213 }
1214
1215 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1216 if (skb->ip_summed != CHECKSUM_PARTIAL)
1217 return 0;
1218
1219 return skb_checksum_setup(skb, recalculate_partial_csum);
1220}
1221
1222static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1223{
1224 u64 now = get_jiffies_64();
1225 u64 next_credit = vif->credit_window_start +
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_eq64(now, next_credit)) {
1234 vif->credit_window_start = 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 vif->credit_window_start = next_credit;
1247
1248 return true;
1249 }
1250
1251 return false;
1252}
1253
1254static void xenvif_tx_build_gops(struct xenvif *vif,
1255 int budget,
1256 unsigned *copy_ops,
1257 unsigned *map_ops)
1258{
1259 struct gnttab_map_grant_ref *gop = vif->tx_map_ops, *request_gop;
1260 struct sk_buff *skb;
1261 int ret;
1262
1263 while (skb_queue_len(&vif->tx_queue) < budget) {
1264 struct xen_netif_tx_request txreq;
1265 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1266 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1267 u16 pending_idx;
1268 RING_IDX idx;
1269 int work_to_do;
1270 unsigned int data_len;
1271 pending_ring_idx_t index;
1272
1273 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1274 XEN_NETIF_TX_RING_SIZE) {
1275 netdev_err(vif->dev,
1276 "Impossible number of requests. "
1277 "req_prod %d, req_cons %d, size %ld\n",
1278 vif->tx.sring->req_prod, vif->tx.req_cons,
1279 XEN_NETIF_TX_RING_SIZE);
1280 xenvif_fatal_tx_err(vif);
1281 break;
1282 }
1283
1284 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);
1285 if (!work_to_do)
1286 break;
1287
1288 idx = vif->tx.req_cons;
1289 rmb(); /* Ensure that we see the request before we copy it. */
1290 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1291
1292 /* Credit-based scheduling. */
1293 if (txreq.size > vif->remaining_credit &&
1294 tx_credit_exceeded(vif, txreq.size))
1295 break;
1296
1297 vif->remaining_credit -= txreq.size;
1298
1299 work_to_do--;
1300 vif->tx.req_cons = ++idx;
1301
1302 memset(extras, 0, sizeof(extras));
1303 if (txreq.flags & XEN_NETTXF_extra_info) {
1304 work_to_do = xenvif_get_extras(vif, extras,
1305 work_to_do);
1306 idx = vif->tx.req_cons;
1307 if (unlikely(work_to_do < 0))
1308 break;
1309 }
1310
1311 ret = xenvif_count_requests(vif, &txreq, txfrags, work_to_do);
1312 if (unlikely(ret < 0))
1313 break;
1314
1315 idx += ret;
1316
1317 if (unlikely(txreq.size < ETH_HLEN)) {
1318 netdev_dbg(vif->dev,
1319 "Bad packet size: %d\n", txreq.size);
1320 xenvif_tx_err(vif, &txreq, idx);
1321 break;
1322 }
1323
1324 /* No crossing a page as the payload mustn't fragment. */
1325 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1326 netdev_err(vif->dev,
1327 "txreq.offset: %x, size: %u, end: %lu\n",
1328 txreq.offset, txreq.size,
1329 (txreq.offset&~PAGE_MASK) + txreq.size);
1330 xenvif_fatal_tx_err(vif);
1331 break;
1332 }
1333
1334 index = pending_index(vif->pending_cons);
1335 pending_idx = vif->pending_ring[index];
1336
1337 data_len = (txreq.size > PKT_PROT_LEN &&
1338 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1339 PKT_PROT_LEN : txreq.size;
1340
1341 skb = xenvif_alloc_skb(data_len);
1342 if (unlikely(skb == NULL)) {
1343 netdev_dbg(vif->dev,
1344 "Can't allocate a skb in start_xmit.\n");
1345 xenvif_tx_err(vif, &txreq, idx);
1346 break;
1347 }
1348
1349 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1350 struct xen_netif_extra_info *gso;
1351 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1352
1353 if (xenvif_set_skb_gso(vif, skb, gso)) {
1354 /* Failure in xenvif_set_skb_gso is fatal. */
1355 kfree_skb(skb);
1356 break;
1357 }
1358 }
1359
1360 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1361
1362 __skb_put(skb, data_len);
1363 vif->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1364 vif->tx_copy_ops[*copy_ops].source.domid = vif->domid;
1365 vif->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1366
1367 vif->tx_copy_ops[*copy_ops].dest.u.gmfn =
1368 virt_to_mfn(skb->data);
1369 vif->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1370 vif->tx_copy_ops[*copy_ops].dest.offset =
1371 offset_in_page(skb->data);
1372
1373 vif->tx_copy_ops[*copy_ops].len = data_len;
1374 vif->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1375
1376 (*copy_ops)++;
1377
1378 skb_shinfo(skb)->nr_frags = ret;
1379 if (data_len < txreq.size) {
1380 skb_shinfo(skb)->nr_frags++;
1381 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1382 pending_idx);
1383 xenvif_tx_create_map_op(vif, pending_idx, &txreq, gop);
1384 gop++;
1385 } else {
1386 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1387 INVALID_PENDING_IDX);
1388 memcpy(&vif->pending_tx_info[pending_idx].req, &txreq,
1389 sizeof(txreq));
1390 }
1391
1392 vif->pending_cons++;
1393
1394 request_gop = xenvif_get_requests(vif, skb, txfrags, gop);
1395 if (request_gop == NULL) {
1396 kfree_skb(skb);
1397 xenvif_tx_err(vif, &txreq, idx);
1398 break;
1399 }
1400 gop = request_gop;
1401
1402 __skb_queue_tail(&vif->tx_queue, skb);
1403
1404 vif->tx.req_cons = idx;
1405
1406 if (((gop-vif->tx_map_ops) >= ARRAY_SIZE(vif->tx_map_ops)) ||
1407 (*copy_ops >= ARRAY_SIZE(vif->tx_copy_ops)))
1408 break;
1409 }
1410
1411 (*map_ops) = gop - vif->tx_map_ops;
1412 return;
1413}
1414
1415/* Consolidate skb with a frag_list into a brand new one with local pages on
1416 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1417 */
1418static int xenvif_handle_frag_list(struct xenvif *vif, struct sk_buff *skb)
1419{
1420 unsigned int offset = skb_headlen(skb);
1421 skb_frag_t frags[MAX_SKB_FRAGS];
1422 int i;
1423 struct ubuf_info *uarg;
1424 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1425
1426 vif->tx_zerocopy_sent += 2;
1427 vif->tx_frag_overflow++;
1428
1429 xenvif_fill_frags(vif, nskb);
1430 /* Subtract frags size, we will correct it later */
1431 skb->truesize -= skb->data_len;
1432 skb->len += nskb->len;
1433 skb->data_len += nskb->len;
1434
1435 /* create a brand new frags array and coalesce there */
1436 for (i = 0; offset < skb->len; i++) {
1437 struct page *page;
1438 unsigned int len;
1439
1440 BUG_ON(i >= MAX_SKB_FRAGS);
1441 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
1442 if (!page) {
1443 int j;
1444 skb->truesize += skb->data_len;
1445 for (j = 0; j < i; j++)
1446 put_page(frags[j].page.p);
1447 return -ENOMEM;
1448 }
1449
1450 if (offset + PAGE_SIZE < skb->len)
1451 len = PAGE_SIZE;
1452 else
1453 len = skb->len - offset;
1454 if (skb_copy_bits(skb, offset, page_address(page), len))
1455 BUG();
1456
1457 offset += len;
1458 frags[i].page.p = page;
1459 frags[i].page_offset = 0;
1460 skb_frag_size_set(&frags[i], len);
1461 }
1462 /* swap out with old one */
1463 memcpy(skb_shinfo(skb)->frags,
1464 frags,
1465 i * sizeof(skb_frag_t));
1466 skb_shinfo(skb)->nr_frags = i;
1467 skb->truesize += i * PAGE_SIZE;
1468
1469 /* remove traces of mapped pages and frag_list */
1470 skb_frag_list_init(skb);
1471 uarg = skb_shinfo(skb)->destructor_arg;
1472 uarg->callback(uarg, true);
1473 skb_shinfo(skb)->destructor_arg = NULL;
1474
1475 skb_shinfo(nskb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
1476 kfree_skb(nskb);
1477
1478 return 0;
1479}
1480
1481static int xenvif_tx_submit(struct xenvif *vif)
1482{
1483 struct gnttab_map_grant_ref *gop_map = vif->tx_map_ops;
1484 struct gnttab_copy *gop_copy = vif->tx_copy_ops;
1485 struct sk_buff *skb;
1486 int work_done = 0;
1487
1488 while ((skb = __skb_dequeue(&vif->tx_queue)) != NULL) {
1489 struct xen_netif_tx_request *txp;
1490 u16 pending_idx;
1491 unsigned data_len;
1492
1493 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1494 txp = &vif->pending_tx_info[pending_idx].req;
1495
1496 /* Check the remap error code. */
1497 if (unlikely(xenvif_tx_check_gop(vif, skb, &gop_map, &gop_copy))) {
1498 skb_shinfo(skb)->nr_frags = 0;
1499 kfree_skb(skb);
1500 continue;
1501 }
1502
1503 data_len = skb->len;
1504 callback_param(vif, pending_idx).ctx = NULL;
1505 if (data_len < txp->size) {
1506 /* Append the packet payload as a fragment. */
1507 txp->offset += data_len;
1508 txp->size -= data_len;
1509 } else {
1510 /* Schedule a response immediately. */
1511 xenvif_idx_release(vif, pending_idx,
1512 XEN_NETIF_RSP_OKAY);
1513 }
1514
1515 if (txp->flags & XEN_NETTXF_csum_blank)
1516 skb->ip_summed = CHECKSUM_PARTIAL;
1517 else if (txp->flags & XEN_NETTXF_data_validated)
1518 skb->ip_summed = CHECKSUM_UNNECESSARY;
1519
1520 xenvif_fill_frags(vif, skb);
1521
1522 if (unlikely(skb_has_frag_list(skb))) {
1523 if (xenvif_handle_frag_list(vif, skb)) {
1524 if (net_ratelimit())
1525 netdev_err(vif->dev,
1526 "Not enough memory to consolidate frag_list!\n");
1527 skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
1528 kfree_skb(skb);
1529 continue;
1530 }
1531 }
1532
1533 if (skb_is_nonlinear(skb) && skb_headlen(skb) < PKT_PROT_LEN) {
1534 int target = min_t(int, skb->len, PKT_PROT_LEN);
1535 __pskb_pull_tail(skb, target - skb_headlen(skb));
1536 }
1537
1538 skb->dev = vif->dev;
1539 skb->protocol = eth_type_trans(skb, skb->dev);
1540 skb_reset_network_header(skb);
1541
1542 if (checksum_setup(vif, skb)) {
1543 netdev_dbg(vif->dev,
1544 "Can't setup checksum in net_tx_action\n");
1545 /* We have to set this flag to trigger the callback */
1546 if (skb_shinfo(skb)->destructor_arg)
1547 skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
1548 kfree_skb(skb);
1549 continue;
1550 }
1551
1552 skb_probe_transport_header(skb, 0);
1553
1554 /* If the packet is GSO then we will have just set up the
1555 * transport header offset in checksum_setup so it's now
1556 * straightforward to calculate gso_segs.
1557 */
1558 if (skb_is_gso(skb)) {
1559 int mss = skb_shinfo(skb)->gso_size;
1560 int hdrlen = skb_transport_header(skb) -
1561 skb_mac_header(skb) +
1562 tcp_hdrlen(skb);
1563
1564 skb_shinfo(skb)->gso_segs =
1565 DIV_ROUND_UP(skb->len - hdrlen, mss);
1566 }
1567
1568 vif->dev->stats.rx_bytes += skb->len;
1569 vif->dev->stats.rx_packets++;
1570
1571 work_done++;
1572
1573 /* Set this flag right before netif_receive_skb, otherwise
1574 * someone might think this packet already left netback, and
1575 * do a skb_copy_ubufs while we are still in control of the
1576 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1577 */
1578 if (skb_shinfo(skb)->destructor_arg) {
1579 skb_shinfo(skb)->tx_flags |= SKBTX_DEV_ZEROCOPY;
1580 vif->tx_zerocopy_sent++;
1581 }
1582
1583 netif_receive_skb(skb);
1584 }
1585
1586 return work_done;
1587}
1588
1589void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1590{
1591 unsigned long flags;
1592 pending_ring_idx_t index;
1593 struct xenvif *vif = ubuf_to_vif(ubuf);
1594
1595 /* This is the only place where we grab this lock, to protect callbacks
1596 * from each other.
1597 */
1598 spin_lock_irqsave(&vif->callback_lock, flags);
1599 do {
1600 u16 pending_idx = ubuf->desc;
1601 ubuf = (struct ubuf_info *) ubuf->ctx;
1602 BUG_ON(vif->dealloc_prod - vif->dealloc_cons >=
1603 MAX_PENDING_REQS);
1604 index = pending_index(vif->dealloc_prod);
1605 vif->dealloc_ring[index] = pending_idx;
1606 /* Sync with xenvif_tx_dealloc_action:
1607 * insert idx then incr producer.
1608 */
1609 smp_wmb();
1610 vif->dealloc_prod++;
1611 } while (ubuf);
1612 wake_up(&vif->dealloc_wq);
1613 spin_unlock_irqrestore(&vif->callback_lock, flags);
1614
1615 if (likely(zerocopy_success))
1616 vif->tx_zerocopy_success++;
1617 else
1618 vif->tx_zerocopy_fail++;
1619}
1620
1621static inline void xenvif_tx_dealloc_action(struct xenvif *vif)
1622{
1623 struct gnttab_unmap_grant_ref *gop;
1624 pending_ring_idx_t dc, dp;
1625 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1626 unsigned int i = 0;
1627
1628 dc = vif->dealloc_cons;
1629 gop = vif->tx_unmap_ops;
1630
1631 /* Free up any grants we have finished using */
1632 do {
1633 dp = vif->dealloc_prod;
1634
1635 /* Ensure we see all indices enqueued by all
1636 * xenvif_zerocopy_callback().
1637 */
1638 smp_rmb();
1639
1640 while (dc != dp) {
1641 BUG_ON(gop - vif->tx_unmap_ops > MAX_PENDING_REQS);
1642 pending_idx =
1643 vif->dealloc_ring[pending_index(dc++)];
1644
1645 pending_idx_release[gop-vif->tx_unmap_ops] =
1646 pending_idx;
1647 vif->pages_to_unmap[gop-vif->tx_unmap_ops] =
1648 vif->mmap_pages[pending_idx];
1649 gnttab_set_unmap_op(gop,
1650 idx_to_kaddr(vif, pending_idx),
1651 GNTMAP_host_map,
1652 vif->grant_tx_handle[pending_idx]);
1653 xenvif_grant_handle_reset(vif, pending_idx);
1654 ++gop;
1655 }
1656
1657 } while (dp != vif->dealloc_prod);
1658
1659 vif->dealloc_cons = dc;
1660
1661 if (gop - vif->tx_unmap_ops > 0) {
1662 int ret;
1663 ret = gnttab_unmap_refs(vif->tx_unmap_ops,
1664 NULL,
1665 vif->pages_to_unmap,
1666 gop - vif->tx_unmap_ops);
1667 if (ret) {
1668 netdev_err(vif->dev, "Unmap fail: nr_ops %tx ret %d\n",
1669 gop - vif->tx_unmap_ops, ret);
1670 for (i = 0; i < gop - vif->tx_unmap_ops; ++i) {
1671 if (gop[i].status != GNTST_okay)
1672 netdev_err(vif->dev,
1673 " host_addr: %llx handle: %x status: %d\n",
1674 gop[i].host_addr,
1675 gop[i].handle,
1676 gop[i].status);
1677 }
1678 BUG();
1679 }
1680 }
1681
1682 for (i = 0; i < gop - vif->tx_unmap_ops; ++i)
1683 xenvif_idx_release(vif, pending_idx_release[i],
1684 XEN_NETIF_RSP_OKAY);
1685}
1686
1687
1688/* Called after netfront has transmitted */
1689int xenvif_tx_action(struct xenvif *vif, int budget)
1690{
1691 unsigned nr_mops, nr_cops = 0;
1692 int work_done, ret;
1693
1694 if (unlikely(!tx_work_todo(vif)))
1695 return 0;
1696
1697 xenvif_tx_build_gops(vif, budget, &nr_cops, &nr_mops);
1698
1699 if (nr_cops == 0)
1700 return 0;
1701
1702 gnttab_batch_copy(vif->tx_copy_ops, nr_cops);
1703 if (nr_mops != 0) {
1704 ret = gnttab_map_refs(vif->tx_map_ops,
1705 NULL,
1706 vif->pages_to_map,
1707 nr_mops);
1708 BUG_ON(ret);
1709 }
1710
1711 work_done = xenvif_tx_submit(vif);
1712
1713 return work_done;
1714}
1715
1716static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
1717 u8 status)
1718{
1719 struct pending_tx_info *pending_tx_info;
1720 pending_ring_idx_t index;
1721 unsigned long flags;
1722
1723 pending_tx_info = &vif->pending_tx_info[pending_idx];
1724 spin_lock_irqsave(&vif->response_lock, flags);
1725 make_tx_response(vif, &pending_tx_info->req, status);
1726 index = pending_index(vif->pending_prod);
1727 vif->pending_ring[index] = pending_idx;
1728 /* TX shouldn't use the index before we give it back here */
1729 mb();
1730 vif->pending_prod++;
1731 spin_unlock_irqrestore(&vif->response_lock, flags);
1732}
1733
1734
1735static void make_tx_response(struct xenvif *vif,
1736 struct xen_netif_tx_request *txp,
1737 s8 st)
1738{
1739 RING_IDX i = vif->tx.rsp_prod_pvt;
1740 struct xen_netif_tx_response *resp;
1741 int notify;
1742
1743 resp = RING_GET_RESPONSE(&vif->tx, i);
1744 resp->id = txp->id;
1745 resp->status = st;
1746
1747 if (txp->flags & XEN_NETTXF_extra_info)
1748 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1749
1750 vif->tx.rsp_prod_pvt = ++i;
1751 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1752 if (notify)
1753 notify_remote_via_irq(vif->tx_irq);
1754}
1755
1756static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1757 u16 id,
1758 s8 st,
1759 u16 offset,
1760 u16 size,
1761 u16 flags)
1762{
1763 RING_IDX i = vif->rx.rsp_prod_pvt;
1764 struct xen_netif_rx_response *resp;
1765
1766 resp = RING_GET_RESPONSE(&vif->rx, i);
1767 resp->offset = offset;
1768 resp->flags = flags;
1769 resp->id = id;
1770 resp->status = (s16)size;
1771 if (st < 0)
1772 resp->status = (s16)st;
1773
1774 vif->rx.rsp_prod_pvt = ++i;
1775
1776 return resp;
1777}
1778
1779void xenvif_idx_unmap(struct xenvif *vif, u16 pending_idx)
1780{
1781 int ret;
1782 struct gnttab_unmap_grant_ref tx_unmap_op;
1783
1784 gnttab_set_unmap_op(&tx_unmap_op,
1785 idx_to_kaddr(vif, pending_idx),
1786 GNTMAP_host_map,
1787 vif->grant_tx_handle[pending_idx]);
1788 xenvif_grant_handle_reset(vif, pending_idx);
1789
1790 ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1791 &vif->mmap_pages[pending_idx], 1);
1792 if (ret) {
1793 netdev_err(vif->dev,
1794 "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: %x status: %d\n",
1795 ret,
1796 pending_idx,
1797 tx_unmap_op.host_addr,
1798 tx_unmap_op.handle,
1799 tx_unmap_op.status);
1800 BUG();
1801 }
1802
1803 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
1804}
1805
1806static inline int rx_work_todo(struct xenvif *vif)
1807{
1808 return (!skb_queue_empty(&vif->rx_queue) &&
1809 xenvif_rx_ring_slots_available(vif, vif->rx_last_skb_slots)) ||
1810 vif->rx_queue_purge;
1811}
1812
1813static inline int tx_work_todo(struct xenvif *vif)
1814{
1815
1816 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->tx)))
1817 return 1;
1818
1819 return 0;
1820}
1821
1822static inline bool tx_dealloc_work_todo(struct xenvif *vif)
1823{
1824 return vif->dealloc_cons != vif->dealloc_prod;
1825}
1826
1827void xenvif_unmap_frontend_rings(struct xenvif *vif)
1828{
1829 if (vif->tx.sring)
1830 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1831 vif->tx.sring);
1832 if (vif->rx.sring)
1833 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1834 vif->rx.sring);
1835}
1836
1837int xenvif_map_frontend_rings(struct xenvif *vif,
1838 grant_ref_t tx_ring_ref,
1839 grant_ref_t rx_ring_ref)
1840{
1841 void *addr;
1842 struct xen_netif_tx_sring *txs;
1843 struct xen_netif_rx_sring *rxs;
1844
1845 int err = -ENOMEM;
1846
1847 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1848 tx_ring_ref, &addr);
1849 if (err)
1850 goto err;
1851
1852 txs = (struct xen_netif_tx_sring *)addr;
1853 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1854
1855 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1856 rx_ring_ref, &addr);
1857 if (err)
1858 goto err;
1859
1860 rxs = (struct xen_netif_rx_sring *)addr;
1861 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1862
1863 return 0;
1864
1865err:
1866 xenvif_unmap_frontend_rings(vif);
1867 return err;
1868}
1869
1870void xenvif_stop_queue(struct xenvif *vif)
1871{
1872 if (!vif->can_queue)
1873 return;
1874
1875 netif_stop_queue(vif->dev);
1876}
1877
1878static void xenvif_start_queue(struct xenvif *vif)
1879{
1880 if (xenvif_schedulable(vif))
1881 netif_wake_queue(vif->dev);
1882}
1883
1884int xenvif_kthread_guest_rx(void *data)
1885{
1886 struct xenvif *vif = data;
1887 struct sk_buff *skb;
1888
1889 while (!kthread_should_stop()) {
1890 wait_event_interruptible(vif->wq,
1891 rx_work_todo(vif) ||
1892 vif->disabled ||
1893 kthread_should_stop());
1894
1895 /* This frontend is found to be rogue, disable it in
1896 * kthread context. Currently this is only set when
1897 * netback finds out frontend sends malformed packet,
1898 * but we cannot disable the interface in softirq
1899 * context so we defer it here.
1900 */
1901 if (unlikely(vif->disabled && netif_carrier_ok(vif->dev)))
1902 xenvif_carrier_off(vif);
1903
1904 if (kthread_should_stop())
1905 break;
1906
1907 if (vif->rx_queue_purge) {
1908 skb_queue_purge(&vif->rx_queue);
1909 vif->rx_queue_purge = false;
1910 }
1911
1912 if (!skb_queue_empty(&vif->rx_queue))
1913 xenvif_rx_action(vif);
1914
1915 if (skb_queue_empty(&vif->rx_queue) &&
1916 netif_queue_stopped(vif->dev)) {
1917 del_timer_sync(&vif->wake_queue);
1918 xenvif_start_queue(vif);
1919 }
1920
1921 cond_resched();
1922 }
1923
1924 /* Bin any remaining skbs */
1925 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL)
1926 dev_kfree_skb(skb);
1927
1928 return 0;
1929}
1930
1931int xenvif_dealloc_kthread(void *data)
1932{
1933 struct xenvif *vif = data;
1934
1935 while (!kthread_should_stop()) {
1936 wait_event_interruptible(vif->dealloc_wq,
1937 tx_dealloc_work_todo(vif) ||
1938 kthread_should_stop());
1939 if (kthread_should_stop())
1940 break;
1941
1942 xenvif_tx_dealloc_action(vif);
1943 cond_resched();
1944 }
1945
1946 /* Unmap anything remaining*/
1947 if (tx_dealloc_work_todo(vif))
1948 xenvif_tx_dealloc_action(vif);
1949
1950 return 0;
1951}
1952
1953static int __init netback_init(void)
1954{
1955 int rc = 0;
1956
1957 if (!xen_domain())
1958 return -ENODEV;
1959
1960 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1961 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1962 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1963 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1964 }
1965
1966 rc = xenvif_xenbus_init();
1967 if (rc)
1968 goto failed_init;
1969
1970 rx_drain_timeout_jiffies = msecs_to_jiffies(rx_drain_timeout_msecs);
1971
1972 return 0;
1973
1974failed_init:
1975 return rc;
1976}
1977
1978module_init(netback_init);
1979
1980static void __exit netback_fini(void)
1981{
1982 xenvif_xenbus_fini();
1983}
1984module_exit(netback_fini);
1985
1986MODULE_LICENSE("Dual BSD/GPL");
1987MODULE_ALIAS("xen-backend:vif");