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
8 * Copyright (c) 2002-2005, K A Fraser
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:
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:
23 * The above copyright notice and this permission notice shall be included in
24 * all copies or substantial portions of the Software.
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
37 #include <linux/kthread.h>
38 #include <linux/if_vlan.h>
39 #include <linux/udp.h>
40 #include <linux/highmem.h>
45 #include <xen/events.h>
46 #include <xen/interface/memory.h>
48 #include <asm/xen/hypercall.h>
49 #include <asm/xen/page.h>
51 /* Provide an option to disable split event channels at load time as
52 * event channels are limited resource. Split event channels are
55 bool separate_tx_rx_irq = 1;
56 module_param(separate_tx_rx_irq, bool, 0644);
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
61 unsigned int rx_drain_timeout_msecs = 10000;
62 module_param(rx_drain_timeout_msecs, uint, 0444);
63 unsigned int rx_drain_timeout_jiffies;
65 unsigned int xenvif_max_queues;
66 module_param_named(max_queues, xenvif_max_queues, uint, 0644);
67 MODULE_PARM_DESC(max_queues,
68 "Maximum number of queues per virtual interface");
71 * This is the maximum slots a skb can have. If a guest sends a skb
72 * which exceeds this limit it is considered malicious.
74 #define FATAL_SKB_SLOTS_DEFAULT 20
75 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
76 module_param(fatal_skb_slots, uint, 0444);
78 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
81 static void make_tx_response(struct xenvif_queue *queue,
82 struct xen_netif_tx_request *txp,
85 static inline int tx_work_todo(struct xenvif_queue *queue);
86 static inline int rx_work_todo(struct xenvif_queue *queue);
88 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
95 static inline unsigned long idx_to_pfn(struct xenvif_queue *queue,
98 return page_to_pfn(queue->mmap_pages[idx]);
101 static inline unsigned long idx_to_kaddr(struct xenvif_queue *queue,
104 return (unsigned long)pfn_to_kaddr(idx_to_pfn(queue, idx));
107 #define callback_param(vif, pending_idx) \
108 (vif->pending_tx_info[pending_idx].callback_struct)
110 /* Find the containing VIF's structure from a pointer in pending_tx_info array
112 static inline struct xenvif_queue *ubuf_to_queue(const struct ubuf_info *ubuf)
114 u16 pending_idx = ubuf->desc;
115 struct pending_tx_info *temp =
116 container_of(ubuf, struct pending_tx_info, callback_struct);
117 return container_of(temp - pending_idx,
122 /* This is a miniumum size for the linear area to avoid lots of
123 * calls to __pskb_pull_tail() as we set up checksum offsets. The
124 * value 128 was chosen as it covers all IPv4 and most likely
127 #define PKT_PROT_LEN 128
129 static u16 frag_get_pending_idx(skb_frag_t *frag)
131 return (u16)frag->page_offset;
134 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
136 frag->page_offset = pending_idx;
139 static inline pending_ring_idx_t pending_index(unsigned i)
141 return i & (MAX_PENDING_REQS-1);
144 bool xenvif_rx_ring_slots_available(struct xenvif_queue *queue, int needed)
149 prod = queue->rx.sring->req_prod;
150 cons = queue->rx.req_cons;
152 if (prod - cons >= needed)
155 queue->rx.sring->req_event = prod + 1;
157 /* Make sure event is visible before we check prod
161 } while (queue->rx.sring->req_prod != prod);
167 * Returns true if we should start a new receive buffer instead of
168 * adding 'size' bytes to a buffer which currently contains 'offset'
171 static bool start_new_rx_buffer(int offset, unsigned long size, int head,
174 /* simple case: we have completely filled the current buffer. */
175 if (offset == MAX_BUFFER_OFFSET)
179 * complex case: start a fresh buffer if the current frag
180 * would overflow the current buffer but only if:
181 * (i) this frag would fit completely in the next buffer
182 * and (ii) there is already some data in the current buffer
183 * and (iii) this is not the head buffer.
184 * and (iv) there is no need to fully utilize the buffers
187 * - (i) stops us splitting a frag into two copies
188 * unless the frag is too large for a single buffer.
189 * - (ii) stops us from leaving a buffer pointlessly empty.
190 * - (iii) stops us leaving the first buffer
191 * empty. Strictly speaking this is already covered
192 * by (ii) but is explicitly checked because
193 * netfront relies on the first buffer being
194 * non-empty and can crash otherwise.
195 * - (iv) is needed for skbs which can use up more than MAX_SKB_FRAGS
198 * This means we will effectively linearise small
199 * frags but do not needlessly split large buffers
200 * into multiple copies tend to give large frags their
201 * own buffers as before.
203 BUG_ON(size > MAX_BUFFER_OFFSET);
204 if ((offset + size > MAX_BUFFER_OFFSET) && offset && !head &&
211 struct netrx_pending_operations {
212 unsigned copy_prod, copy_cons;
213 unsigned meta_prod, meta_cons;
214 struct gnttab_copy *copy;
215 struct xenvif_rx_meta *meta;
217 grant_ref_t copy_gref;
220 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif_queue *queue,
221 struct netrx_pending_operations *npo)
223 struct xenvif_rx_meta *meta;
224 struct xen_netif_rx_request *req;
226 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
228 meta = npo->meta + npo->meta_prod++;
229 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
235 npo->copy_gref = req->gref;
240 struct xenvif_rx_cb {
245 #define XENVIF_RX_CB(skb) ((struct xenvif_rx_cb *)(skb)->cb)
248 * Set up the grant operations for this fragment. If it's a flipping
249 * interface, we also set up the unmap request from here.
251 static void xenvif_gop_frag_copy(struct xenvif_queue *queue, struct sk_buff *skb,
252 struct netrx_pending_operations *npo,
253 struct page *page, unsigned long size,
254 unsigned long offset, int *head,
255 struct xenvif_queue *foreign_queue,
256 grant_ref_t foreign_gref)
258 struct gnttab_copy *copy_gop;
259 struct xenvif_rx_meta *meta;
261 int gso_type = XEN_NETIF_GSO_TYPE_NONE;
263 /* Data must not cross a page boundary. */
264 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
266 meta = npo->meta + npo->meta_prod - 1;
268 /* Skip unused frames from start of page */
269 page += offset >> PAGE_SHIFT;
270 offset &= ~PAGE_MASK;
273 BUG_ON(offset >= PAGE_SIZE);
274 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
276 bytes = PAGE_SIZE - offset;
281 if (start_new_rx_buffer(npo->copy_off,
284 XENVIF_RX_CB(skb)->full_coalesce)) {
286 * Netfront requires there to be some data in the head
291 meta = get_next_rx_buffer(queue, npo);
294 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
295 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
297 copy_gop = npo->copy + npo->copy_prod++;
298 copy_gop->flags = GNTCOPY_dest_gref;
299 copy_gop->len = bytes;
302 copy_gop->source.domid = foreign_queue->vif->domid;
303 copy_gop->source.u.ref = foreign_gref;
304 copy_gop->flags |= GNTCOPY_source_gref;
306 copy_gop->source.domid = DOMID_SELF;
307 copy_gop->source.u.gmfn =
308 virt_to_mfn(page_address(page));
310 copy_gop->source.offset = offset;
312 copy_gop->dest.domid = queue->vif->domid;
313 copy_gop->dest.offset = npo->copy_off;
314 copy_gop->dest.u.ref = npo->copy_gref;
316 npo->copy_off += bytes;
323 if (offset == PAGE_SIZE && size) {
324 BUG_ON(!PageCompound(page));
329 /* Leave a gap for the GSO descriptor. */
330 if (skb_is_gso(skb)) {
331 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
332 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
333 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
334 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
337 if (*head && ((1 << gso_type) & queue->vif->gso_mask))
338 queue->rx.req_cons++;
340 *head = 0; /* There must be something in this buffer now. */
346 * Find the grant ref for a given frag in a chain of struct ubuf_info's
347 * skb: the skb itself
348 * i: the frag's number
349 * ubuf: a pointer to an element in the chain. It should not be NULL
351 * Returns a pointer to the element in the chain where the page were found. If
352 * not found, returns NULL.
353 * See the definition of callback_struct in common.h for more details about
356 static const struct ubuf_info *xenvif_find_gref(const struct sk_buff *const skb,
358 const struct ubuf_info *ubuf)
360 struct xenvif_queue *foreign_queue = ubuf_to_queue(ubuf);
363 u16 pending_idx = ubuf->desc;
365 if (skb_shinfo(skb)->frags[i].page.p ==
366 foreign_queue->mmap_pages[pending_idx])
368 ubuf = (struct ubuf_info *) ubuf->ctx;
375 * Prepare an SKB to be transmitted to the frontend.
377 * This function is responsible for allocating grant operations, meta
380 * It returns the number of meta structures consumed. The number of
381 * ring slots used is always equal to the number of meta slots used
382 * plus the number of GSO descriptors used. Currently, we use either
383 * zero GSO descriptors (for non-GSO packets) or one descriptor (for
384 * frontend-side LRO).
386 static int xenvif_gop_skb(struct sk_buff *skb,
387 struct netrx_pending_operations *npo,
388 struct xenvif_queue *queue)
390 struct xenvif *vif = netdev_priv(skb->dev);
391 int nr_frags = skb_shinfo(skb)->nr_frags;
393 struct xen_netif_rx_request *req;
394 struct xenvif_rx_meta *meta;
399 const struct ubuf_info *ubuf = skb_shinfo(skb)->destructor_arg;
400 const struct ubuf_info *const head_ubuf = ubuf;
402 old_meta_prod = npo->meta_prod;
404 gso_type = XEN_NETIF_GSO_TYPE_NONE;
405 if (skb_is_gso(skb)) {
406 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
407 gso_type = XEN_NETIF_GSO_TYPE_TCPV4;
408 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
409 gso_type = XEN_NETIF_GSO_TYPE_TCPV6;
412 /* Set up a GSO prefix descriptor, if necessary */
413 if ((1 << gso_type) & vif->gso_prefix_mask) {
414 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
415 meta = npo->meta + npo->meta_prod++;
416 meta->gso_type = gso_type;
417 meta->gso_size = skb_shinfo(skb)->gso_size;
422 req = RING_GET_REQUEST(&queue->rx, queue->rx.req_cons++);
423 meta = npo->meta + npo->meta_prod++;
425 if ((1 << gso_type) & vif->gso_mask) {
426 meta->gso_type = gso_type;
427 meta->gso_size = skb_shinfo(skb)->gso_size;
429 meta->gso_type = XEN_NETIF_GSO_TYPE_NONE;
436 npo->copy_gref = req->gref;
439 while (data < skb_tail_pointer(skb)) {
440 unsigned int offset = offset_in_page(data);
441 unsigned int len = PAGE_SIZE - offset;
443 if (data + len > skb_tail_pointer(skb))
444 len = skb_tail_pointer(skb) - data;
446 xenvif_gop_frag_copy(queue, skb, npo,
447 virt_to_page(data), len, offset, &head,
453 for (i = 0; i < nr_frags; i++) {
454 /* This variable also signals whether foreign_gref has a real
457 struct xenvif_queue *foreign_queue = NULL;
458 grant_ref_t foreign_gref;
460 if ((skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY) &&
461 (ubuf->callback == &xenvif_zerocopy_callback)) {
462 const struct ubuf_info *const startpoint = ubuf;
464 /* Ideally ubuf points to the chain element which
465 * belongs to this frag. Or if frags were removed from
466 * the beginning, then shortly before it.
468 ubuf = xenvif_find_gref(skb, i, ubuf);
470 /* Try again from the beginning of the list, if we
471 * haven't tried from there. This only makes sense in
472 * the unlikely event of reordering the original frags.
473 * For injected local pages it's an unnecessary second
476 if (unlikely(!ubuf) && startpoint != head_ubuf)
477 ubuf = xenvif_find_gref(skb, i, head_ubuf);
480 u16 pending_idx = ubuf->desc;
482 foreign_queue = ubuf_to_queue(ubuf);
484 foreign_queue->pending_tx_info[pending_idx].req.gref;
485 /* Just a safety measure. If this was the last
486 * element on the list, the for loop will
487 * iterate again if a local page were added to
488 * the end. Using head_ubuf here prevents the
489 * second search on the chain. Or the original
490 * frags changed order, but that's less likely.
491 * In any way, ubuf shouldn't be NULL.
494 (struct ubuf_info *) ubuf->ctx :
497 /* This frag was a local page, added to the
498 * array after the skb left netback.
502 xenvif_gop_frag_copy(queue, skb, npo,
503 skb_frag_page(&skb_shinfo(skb)->frags[i]),
504 skb_frag_size(&skb_shinfo(skb)->frags[i]),
505 skb_shinfo(skb)->frags[i].page_offset,
508 foreign_queue ? foreign_gref : UINT_MAX);
511 return npo->meta_prod - old_meta_prod;
515 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
516 * used to set up the operations on the top of
517 * netrx_pending_operations, which have since been done. Check that
518 * they didn't give any errors and advance over them.
520 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
521 struct netrx_pending_operations *npo)
523 struct gnttab_copy *copy_op;
524 int status = XEN_NETIF_RSP_OKAY;
527 for (i = 0; i < nr_meta_slots; i++) {
528 copy_op = npo->copy + npo->copy_cons++;
529 if (copy_op->status != GNTST_okay) {
531 "Bad status %d from copy to DOM%d.\n",
532 copy_op->status, vif->domid);
533 status = XEN_NETIF_RSP_ERROR;
540 static void xenvif_add_frag_responses(struct xenvif_queue *queue, int status,
541 struct xenvif_rx_meta *meta,
545 unsigned long offset;
547 /* No fragments used */
548 if (nr_meta_slots <= 1)
553 for (i = 0; i < nr_meta_slots; i++) {
555 if (i == nr_meta_slots - 1)
558 flags = XEN_NETRXF_more_data;
561 make_rx_response(queue, meta[i].id, status, offset,
562 meta[i].size, flags);
566 void xenvif_kick_thread(struct xenvif_queue *queue)
571 static void xenvif_rx_action(struct xenvif_queue *queue)
575 struct xen_netif_rx_response *resp;
576 struct sk_buff_head rxq;
580 unsigned long offset;
581 bool need_to_notify = false;
583 struct netrx_pending_operations npo = {
584 .copy = queue->grant_copy_op,
588 skb_queue_head_init(&rxq);
590 while ((skb = skb_dequeue(&queue->rx_queue)) != NULL) {
591 RING_IDX max_slots_needed;
592 RING_IDX old_req_cons;
593 RING_IDX ring_slots_used;
596 /* We need a cheap worse case estimate for the number of
600 max_slots_needed = DIV_ROUND_UP(offset_in_page(skb->data) +
603 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
607 size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
608 offset = skb_shinfo(skb)->frags[i].page_offset;
610 /* For a worse-case estimate we need to factor in
611 * the fragment page offset as this will affect the
612 * number of times xenvif_gop_frag_copy() will
613 * call start_new_rx_buffer().
615 max_slots_needed += DIV_ROUND_UP(offset + size,
619 /* To avoid the estimate becoming too pessimal for some
620 * frontends that limit posted rx requests, cap the estimate
621 * at MAX_SKB_FRAGS. In this case netback will fully coalesce
622 * the skb into the provided slots.
624 if (max_slots_needed > MAX_SKB_FRAGS) {
625 max_slots_needed = MAX_SKB_FRAGS;
626 XENVIF_RX_CB(skb)->full_coalesce = true;
628 XENVIF_RX_CB(skb)->full_coalesce = false;
631 /* We may need one more slot for GSO metadata */
632 if (skb_is_gso(skb) &&
633 (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4 ||
634 skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6))
637 /* If the skb may not fit then bail out now */
638 if (!xenvif_rx_ring_slots_available(queue, max_slots_needed)) {
639 skb_queue_head(&queue->rx_queue, skb);
640 need_to_notify = true;
641 queue->rx_last_skb_slots = max_slots_needed;
644 queue->rx_last_skb_slots = 0;
646 old_req_cons = queue->rx.req_cons;
647 XENVIF_RX_CB(skb)->meta_slots_used = xenvif_gop_skb(skb, &npo, queue);
648 ring_slots_used = queue->rx.req_cons - old_req_cons;
650 BUG_ON(ring_slots_used > max_slots_needed);
652 __skb_queue_tail(&rxq, skb);
655 BUG_ON(npo.meta_prod > ARRAY_SIZE(queue->meta));
660 BUG_ON(npo.copy_prod > MAX_GRANT_COPY_OPS);
661 gnttab_batch_copy(queue->grant_copy_op, npo.copy_prod);
663 while ((skb = __skb_dequeue(&rxq)) != NULL) {
665 if ((1 << queue->meta[npo.meta_cons].gso_type) &
666 queue->vif->gso_prefix_mask) {
667 resp = RING_GET_RESPONSE(&queue->rx,
668 queue->rx.rsp_prod_pvt++);
670 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
672 resp->offset = queue->meta[npo.meta_cons].gso_size;
673 resp->id = queue->meta[npo.meta_cons].id;
674 resp->status = XENVIF_RX_CB(skb)->meta_slots_used;
677 XENVIF_RX_CB(skb)->meta_slots_used--;
681 queue->stats.tx_bytes += skb->len;
682 queue->stats.tx_packets++;
684 status = xenvif_check_gop(queue->vif,
685 XENVIF_RX_CB(skb)->meta_slots_used,
688 if (XENVIF_RX_CB(skb)->meta_slots_used == 1)
691 flags = XEN_NETRXF_more_data;
693 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
694 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
695 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
696 /* remote but checksummed. */
697 flags |= XEN_NETRXF_data_validated;
700 resp = make_rx_response(queue, queue->meta[npo.meta_cons].id,
702 queue->meta[npo.meta_cons].size,
705 if ((1 << queue->meta[npo.meta_cons].gso_type) &
706 queue->vif->gso_mask) {
707 struct xen_netif_extra_info *gso =
708 (struct xen_netif_extra_info *)
709 RING_GET_RESPONSE(&queue->rx,
710 queue->rx.rsp_prod_pvt++);
712 resp->flags |= XEN_NETRXF_extra_info;
714 gso->u.gso.type = queue->meta[npo.meta_cons].gso_type;
715 gso->u.gso.size = queue->meta[npo.meta_cons].gso_size;
717 gso->u.gso.features = 0;
719 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
723 xenvif_add_frag_responses(queue, status,
724 queue->meta + npo.meta_cons + 1,
725 XENVIF_RX_CB(skb)->meta_slots_used);
727 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, ret);
729 need_to_notify |= !!ret;
731 npo.meta_cons += XENVIF_RX_CB(skb)->meta_slots_used;
737 notify_remote_via_irq(queue->rx_irq);
740 void xenvif_napi_schedule_or_enable_events(struct xenvif_queue *queue)
744 RING_FINAL_CHECK_FOR_REQUESTS(&queue->tx, more_to_do);
747 napi_schedule(&queue->napi);
750 static void tx_add_credit(struct xenvif_queue *queue)
752 unsigned long max_burst, max_credit;
755 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
756 * Otherwise the interface can seize up due to insufficient credit.
758 max_burst = RING_GET_REQUEST(&queue->tx, queue->tx.req_cons)->size;
759 max_burst = min(max_burst, 131072UL);
760 max_burst = max(max_burst, queue->credit_bytes);
762 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
763 max_credit = queue->remaining_credit + queue->credit_bytes;
764 if (max_credit < queue->remaining_credit)
765 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
767 queue->remaining_credit = min(max_credit, max_burst);
770 static void tx_credit_callback(unsigned long data)
772 struct xenvif_queue *queue = (struct xenvif_queue *)data;
773 tx_add_credit(queue);
774 xenvif_napi_schedule_or_enable_events(queue);
777 static void xenvif_tx_err(struct xenvif_queue *queue,
778 struct xen_netif_tx_request *txp, RING_IDX end)
780 RING_IDX cons = queue->tx.req_cons;
784 spin_lock_irqsave(&queue->response_lock, flags);
785 make_tx_response(queue, txp, XEN_NETIF_RSP_ERROR);
786 spin_unlock_irqrestore(&queue->response_lock, flags);
789 txp = RING_GET_REQUEST(&queue->tx, cons++);
791 queue->tx.req_cons = cons;
794 static void xenvif_fatal_tx_err(struct xenvif *vif)
796 netdev_err(vif->dev, "fatal error; disabling device\n");
797 vif->disabled = true;
798 /* Disable the vif from queue 0's kthread */
800 xenvif_kick_thread(&vif->queues[0]);
803 static int xenvif_count_requests(struct xenvif_queue *queue,
804 struct xen_netif_tx_request *first,
805 struct xen_netif_tx_request *txp,
808 RING_IDX cons = queue->tx.req_cons;
813 if (!(first->flags & XEN_NETTXF_more_data))
817 struct xen_netif_tx_request dropped_tx = { 0 };
819 if (slots >= work_to_do) {
820 netdev_err(queue->vif->dev,
821 "Asked for %d slots but exceeds this limit\n",
823 xenvif_fatal_tx_err(queue->vif);
827 /* This guest is really using too many slots and
828 * considered malicious.
830 if (unlikely(slots >= fatal_skb_slots)) {
831 netdev_err(queue->vif->dev,
832 "Malicious frontend using %d slots, threshold %u\n",
833 slots, fatal_skb_slots);
834 xenvif_fatal_tx_err(queue->vif);
838 /* Xen network protocol had implicit dependency on
839 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
840 * the historical MAX_SKB_FRAGS value 18 to honor the
841 * same behavior as before. Any packet using more than
842 * 18 slots but less than fatal_skb_slots slots is
845 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
847 netdev_dbg(queue->vif->dev,
848 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
849 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
856 memcpy(txp, RING_GET_REQUEST(&queue->tx, cons + slots),
859 /* If the guest submitted a frame >= 64 KiB then
860 * first->size overflowed and following slots will
861 * appear to be larger than the frame.
863 * This cannot be fatal error as there are buggy
864 * frontends that do this.
866 * Consume all slots and drop the packet.
868 if (!drop_err && txp->size > first->size) {
870 netdev_dbg(queue->vif->dev,
871 "Invalid tx request, slot size %u > remaining size %u\n",
872 txp->size, first->size);
876 first->size -= txp->size;
879 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
880 netdev_err(queue->vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
881 txp->offset, txp->size);
882 xenvif_fatal_tx_err(queue->vif);
886 more_data = txp->flags & XEN_NETTXF_more_data;
894 xenvif_tx_err(queue, first, cons + slots);
902 struct xenvif_tx_cb {
906 #define XENVIF_TX_CB(skb) ((struct xenvif_tx_cb *)(skb)->cb)
908 static inline void xenvif_tx_create_map_op(struct xenvif_queue *queue,
910 struct xen_netif_tx_request *txp,
911 struct gnttab_map_grant_ref *mop)
913 queue->pages_to_map[mop-queue->tx_map_ops] = queue->mmap_pages[pending_idx];
914 gnttab_set_map_op(mop, idx_to_kaddr(queue, pending_idx),
915 GNTMAP_host_map | GNTMAP_readonly,
916 txp->gref, queue->vif->domid);
918 memcpy(&queue->pending_tx_info[pending_idx].req, txp,
922 static inline struct sk_buff *xenvif_alloc_skb(unsigned int size)
924 struct sk_buff *skb =
925 alloc_skb(size + NET_SKB_PAD + NET_IP_ALIGN,
926 GFP_ATOMIC | __GFP_NOWARN);
927 if (unlikely(skb == NULL))
930 /* Packets passed to netif_rx() must have some headroom. */
931 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
933 /* Initialize it here to avoid later surprises */
934 skb_shinfo(skb)->destructor_arg = NULL;
939 static struct gnttab_map_grant_ref *xenvif_get_requests(struct xenvif_queue *queue,
941 struct xen_netif_tx_request *txp,
942 struct gnttab_map_grant_ref *gop)
944 struct skb_shared_info *shinfo = skb_shinfo(skb);
945 skb_frag_t *frags = shinfo->frags;
946 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
948 pending_ring_idx_t index;
949 unsigned int nr_slots, frag_overflow = 0;
951 /* At this point shinfo->nr_frags is in fact the number of
952 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
954 if (shinfo->nr_frags > MAX_SKB_FRAGS) {
955 frag_overflow = shinfo->nr_frags - MAX_SKB_FRAGS;
956 BUG_ON(frag_overflow > MAX_SKB_FRAGS);
957 shinfo->nr_frags = MAX_SKB_FRAGS;
959 nr_slots = shinfo->nr_frags;
961 /* Skip first skb fragment if it is on same page as header fragment. */
962 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
964 for (shinfo->nr_frags = start; shinfo->nr_frags < nr_slots;
965 shinfo->nr_frags++, txp++, gop++) {
966 index = pending_index(queue->pending_cons++);
967 pending_idx = queue->pending_ring[index];
968 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
969 frag_set_pending_idx(&frags[shinfo->nr_frags], pending_idx);
973 struct sk_buff *nskb = xenvif_alloc_skb(0);
974 if (unlikely(nskb == NULL)) {
976 netdev_err(queue->vif->dev,
977 "Can't allocate the frag_list skb.\n");
981 shinfo = skb_shinfo(nskb);
982 frags = shinfo->frags;
984 for (shinfo->nr_frags = 0; shinfo->nr_frags < frag_overflow;
985 shinfo->nr_frags++, txp++, gop++) {
986 index = pending_index(queue->pending_cons++);
987 pending_idx = queue->pending_ring[index];
988 xenvif_tx_create_map_op(queue, pending_idx, txp, gop);
989 frag_set_pending_idx(&frags[shinfo->nr_frags],
993 skb_shinfo(skb)->frag_list = nskb;
999 static inline void xenvif_grant_handle_set(struct xenvif_queue *queue,
1001 grant_handle_t handle)
1003 if (unlikely(queue->grant_tx_handle[pending_idx] !=
1004 NETBACK_INVALID_HANDLE)) {
1005 netdev_err(queue->vif->dev,
1006 "Trying to overwrite active handle! pending_idx: %x\n",
1010 queue->grant_tx_handle[pending_idx] = handle;
1013 static inline void xenvif_grant_handle_reset(struct xenvif_queue *queue,
1016 if (unlikely(queue->grant_tx_handle[pending_idx] ==
1017 NETBACK_INVALID_HANDLE)) {
1018 netdev_err(queue->vif->dev,
1019 "Trying to unmap invalid handle! pending_idx: %x\n",
1023 queue->grant_tx_handle[pending_idx] = NETBACK_INVALID_HANDLE;
1026 static int xenvif_tx_check_gop(struct xenvif_queue *queue,
1027 struct sk_buff *skb,
1028 struct gnttab_map_grant_ref **gopp_map,
1029 struct gnttab_copy **gopp_copy)
1031 struct gnttab_map_grant_ref *gop_map = *gopp_map;
1032 u16 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1033 /* This always points to the shinfo of the skb being checked, which
1034 * could be either the first or the one on the frag_list
1036 struct skb_shared_info *shinfo = skb_shinfo(skb);
1037 /* If this is non-NULL, we are currently checking the frag_list skb, and
1038 * this points to the shinfo of the first one
1040 struct skb_shared_info *first_shinfo = NULL;
1041 int nr_frags = shinfo->nr_frags;
1042 const bool sharedslot = nr_frags &&
1043 frag_get_pending_idx(&shinfo->frags[0]) == pending_idx;
1046 /* Check status of header. */
1047 err = (*gopp_copy)->status;
1048 if (unlikely(err)) {
1049 if (net_ratelimit())
1050 netdev_dbg(queue->vif->dev,
1051 "Grant copy of header failed! status: %d pending_idx: %u ref: %u\n",
1052 (*gopp_copy)->status,
1054 (*gopp_copy)->source.u.ref);
1055 /* The first frag might still have this slot mapped */
1057 xenvif_idx_release(queue, pending_idx,
1058 XEN_NETIF_RSP_ERROR);
1063 for (i = 0; i < nr_frags; i++, gop_map++) {
1066 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
1068 /* Check error status: if okay then remember grant handle. */
1069 newerr = gop_map->status;
1071 if (likely(!newerr)) {
1072 xenvif_grant_handle_set(queue,
1075 /* Had a previous error? Invalidate this fragment. */
1076 if (unlikely(err)) {
1077 xenvif_idx_unmap(queue, pending_idx);
1078 /* If the mapping of the first frag was OK, but
1079 * the header's copy failed, and they are
1080 * sharing a slot, send an error
1082 if (i == 0 && sharedslot)
1083 xenvif_idx_release(queue, pending_idx,
1084 XEN_NETIF_RSP_ERROR);
1086 xenvif_idx_release(queue, pending_idx,
1087 XEN_NETIF_RSP_OKAY);
1092 /* Error on this fragment: respond to client with an error. */
1093 if (net_ratelimit())
1094 netdev_dbg(queue->vif->dev,
1095 "Grant map of %d. frag failed! status: %d pending_idx: %u ref: %u\n",
1101 xenvif_idx_release(queue, pending_idx, XEN_NETIF_RSP_ERROR);
1103 /* Not the first error? Preceding frags already invalidated. */
1107 /* First error: if the header haven't shared a slot with the
1108 * first frag, release it as well.
1111 xenvif_idx_release(queue,
1112 XENVIF_TX_CB(skb)->pending_idx,
1113 XEN_NETIF_RSP_OKAY);
1115 /* Invalidate preceding fragments of this skb. */
1116 for (j = 0; j < i; j++) {
1117 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
1118 xenvif_idx_unmap(queue, pending_idx);
1119 xenvif_idx_release(queue, pending_idx,
1120 XEN_NETIF_RSP_OKAY);
1123 /* And if we found the error while checking the frag_list, unmap
1124 * the first skb's frags
1127 for (j = 0; j < first_shinfo->nr_frags; j++) {
1128 pending_idx = frag_get_pending_idx(&first_shinfo->frags[j]);
1129 xenvif_idx_unmap(queue, pending_idx);
1130 xenvif_idx_release(queue, pending_idx,
1131 XEN_NETIF_RSP_OKAY);
1135 /* Remember the error: invalidate all subsequent fragments. */
1139 if (skb_has_frag_list(skb) && !first_shinfo) {
1140 first_shinfo = skb_shinfo(skb);
1141 shinfo = skb_shinfo(skb_shinfo(skb)->frag_list);
1142 nr_frags = shinfo->nr_frags;
1147 *gopp_map = gop_map;
1151 static void xenvif_fill_frags(struct xenvif_queue *queue, struct sk_buff *skb)
1153 struct skb_shared_info *shinfo = skb_shinfo(skb);
1154 int nr_frags = shinfo->nr_frags;
1156 u16 prev_pending_idx = INVALID_PENDING_IDX;
1158 for (i = 0; i < nr_frags; i++) {
1159 skb_frag_t *frag = shinfo->frags + i;
1160 struct xen_netif_tx_request *txp;
1164 pending_idx = frag_get_pending_idx(frag);
1166 /* If this is not the first frag, chain it to the previous*/
1167 if (prev_pending_idx == INVALID_PENDING_IDX)
1168 skb_shinfo(skb)->destructor_arg =
1169 &callback_param(queue, pending_idx);
1171 callback_param(queue, prev_pending_idx).ctx =
1172 &callback_param(queue, pending_idx);
1174 callback_param(queue, pending_idx).ctx = NULL;
1175 prev_pending_idx = pending_idx;
1177 txp = &queue->pending_tx_info[pending_idx].req;
1178 page = virt_to_page(idx_to_kaddr(queue, pending_idx));
1179 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1180 skb->len += txp->size;
1181 skb->data_len += txp->size;
1182 skb->truesize += txp->size;
1184 /* Take an extra reference to offset network stack's put_page */
1185 get_page(queue->mmap_pages[pending_idx]);
1187 /* FIXME: __skb_fill_page_desc set this to true because page->pfmemalloc
1188 * overlaps with "index", and "mapping" is not set. I think mapping
1189 * should be set. If delivered to local stack, it would drop this
1190 * skb in sk_filter unless the socket has the right to use it.
1192 skb->pfmemalloc = false;
1195 static int xenvif_get_extras(struct xenvif_queue *queue,
1196 struct xen_netif_extra_info *extras,
1199 struct xen_netif_extra_info extra;
1200 RING_IDX cons = queue->tx.req_cons;
1203 if (unlikely(work_to_do-- <= 0)) {
1204 netdev_err(queue->vif->dev, "Missing extra info\n");
1205 xenvif_fatal_tx_err(queue->vif);
1209 memcpy(&extra, RING_GET_REQUEST(&queue->tx, cons),
1211 if (unlikely(!extra.type ||
1212 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1213 queue->tx.req_cons = ++cons;
1214 netdev_err(queue->vif->dev,
1215 "Invalid extra type: %d\n", extra.type);
1216 xenvif_fatal_tx_err(queue->vif);
1220 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1221 queue->tx.req_cons = ++cons;
1222 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1227 static int xenvif_set_skb_gso(struct xenvif *vif,
1228 struct sk_buff *skb,
1229 struct xen_netif_extra_info *gso)
1231 if (!gso->u.gso.size) {
1232 netdev_err(vif->dev, "GSO size must not be zero.\n");
1233 xenvif_fatal_tx_err(vif);
1237 switch (gso->u.gso.type) {
1238 case XEN_NETIF_GSO_TYPE_TCPV4:
1239 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1241 case XEN_NETIF_GSO_TYPE_TCPV6:
1242 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
1245 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1246 xenvif_fatal_tx_err(vif);
1250 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1251 /* gso_segs will be calculated later */
1256 static int checksum_setup(struct xenvif_queue *queue, struct sk_buff *skb)
1258 bool recalculate_partial_csum = false;
1260 /* A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1261 * peers can fail to set NETRXF_csum_blank when sending a GSO
1262 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1263 * recalculate the partial checksum.
1265 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1266 queue->stats.rx_gso_checksum_fixup++;
1267 skb->ip_summed = CHECKSUM_PARTIAL;
1268 recalculate_partial_csum = true;
1271 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1272 if (skb->ip_summed != CHECKSUM_PARTIAL)
1275 return skb_checksum_setup(skb, recalculate_partial_csum);
1278 static bool tx_credit_exceeded(struct xenvif_queue *queue, unsigned size)
1280 u64 now = get_jiffies_64();
1281 u64 next_credit = queue->credit_window_start +
1282 msecs_to_jiffies(queue->credit_usec / 1000);
1284 /* Timer could already be pending in rare cases. */
1285 if (timer_pending(&queue->credit_timeout))
1288 /* Passed the point where we can replenish credit? */
1289 if (time_after_eq64(now, next_credit)) {
1290 queue->credit_window_start = now;
1291 tx_add_credit(queue);
1294 /* Still too big to send right now? Set a callback. */
1295 if (size > queue->remaining_credit) {
1296 queue->credit_timeout.data =
1297 (unsigned long)queue;
1298 queue->credit_timeout.function =
1300 mod_timer(&queue->credit_timeout,
1302 queue->credit_window_start = next_credit;
1310 static void xenvif_tx_build_gops(struct xenvif_queue *queue,
1315 struct gnttab_map_grant_ref *gop = queue->tx_map_ops, *request_gop;
1316 struct sk_buff *skb;
1319 while (skb_queue_len(&queue->tx_queue) < budget) {
1320 struct xen_netif_tx_request txreq;
1321 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1322 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1326 unsigned int data_len;
1327 pending_ring_idx_t index;
1329 if (queue->tx.sring->req_prod - queue->tx.req_cons >
1330 XEN_NETIF_TX_RING_SIZE) {
1331 netdev_err(queue->vif->dev,
1332 "Impossible number of requests. "
1333 "req_prod %d, req_cons %d, size %ld\n",
1334 queue->tx.sring->req_prod, queue->tx.req_cons,
1335 XEN_NETIF_TX_RING_SIZE);
1336 xenvif_fatal_tx_err(queue->vif);
1340 work_to_do = RING_HAS_UNCONSUMED_REQUESTS(&queue->tx);
1344 idx = queue->tx.req_cons;
1345 rmb(); /* Ensure that we see the request before we copy it. */
1346 memcpy(&txreq, RING_GET_REQUEST(&queue->tx, idx), sizeof(txreq));
1348 /* Credit-based scheduling. */
1349 if (txreq.size > queue->remaining_credit &&
1350 tx_credit_exceeded(queue, txreq.size))
1353 queue->remaining_credit -= txreq.size;
1356 queue->tx.req_cons = ++idx;
1358 memset(extras, 0, sizeof(extras));
1359 if (txreq.flags & XEN_NETTXF_extra_info) {
1360 work_to_do = xenvif_get_extras(queue, extras,
1362 idx = queue->tx.req_cons;
1363 if (unlikely(work_to_do < 0))
1367 ret = xenvif_count_requests(queue, &txreq, txfrags, work_to_do);
1368 if (unlikely(ret < 0))
1373 if (unlikely(txreq.size < ETH_HLEN)) {
1374 netdev_dbg(queue->vif->dev,
1375 "Bad packet size: %d\n", txreq.size);
1376 xenvif_tx_err(queue, &txreq, idx);
1380 /* No crossing a page as the payload mustn't fragment. */
1381 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1382 netdev_err(queue->vif->dev,
1383 "txreq.offset: %x, size: %u, end: %lu\n",
1384 txreq.offset, txreq.size,
1385 (txreq.offset&~PAGE_MASK) + txreq.size);
1386 xenvif_fatal_tx_err(queue->vif);
1390 index = pending_index(queue->pending_cons);
1391 pending_idx = queue->pending_ring[index];
1393 data_len = (txreq.size > PKT_PROT_LEN &&
1394 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1395 PKT_PROT_LEN : txreq.size;
1397 skb = xenvif_alloc_skb(data_len);
1398 if (unlikely(skb == NULL)) {
1399 netdev_dbg(queue->vif->dev,
1400 "Can't allocate a skb in start_xmit.\n");
1401 xenvif_tx_err(queue, &txreq, idx);
1405 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1406 struct xen_netif_extra_info *gso;
1407 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1409 if (xenvif_set_skb_gso(queue->vif, skb, gso)) {
1410 /* Failure in xenvif_set_skb_gso is fatal. */
1416 XENVIF_TX_CB(skb)->pending_idx = pending_idx;
1418 __skb_put(skb, data_len);
1419 queue->tx_copy_ops[*copy_ops].source.u.ref = txreq.gref;
1420 queue->tx_copy_ops[*copy_ops].source.domid = queue->vif->domid;
1421 queue->tx_copy_ops[*copy_ops].source.offset = txreq.offset;
1423 queue->tx_copy_ops[*copy_ops].dest.u.gmfn =
1424 virt_to_mfn(skb->data);
1425 queue->tx_copy_ops[*copy_ops].dest.domid = DOMID_SELF;
1426 queue->tx_copy_ops[*copy_ops].dest.offset =
1427 offset_in_page(skb->data);
1429 queue->tx_copy_ops[*copy_ops].len = data_len;
1430 queue->tx_copy_ops[*copy_ops].flags = GNTCOPY_source_gref;
1434 skb_shinfo(skb)->nr_frags = ret;
1435 if (data_len < txreq.size) {
1436 skb_shinfo(skb)->nr_frags++;
1437 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1439 xenvif_tx_create_map_op(queue, pending_idx, &txreq, gop);
1442 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1443 INVALID_PENDING_IDX);
1444 memcpy(&queue->pending_tx_info[pending_idx].req, &txreq,
1448 queue->pending_cons++;
1450 request_gop = xenvif_get_requests(queue, skb, txfrags, gop);
1451 if (request_gop == NULL) {
1453 xenvif_tx_err(queue, &txreq, idx);
1458 __skb_queue_tail(&queue->tx_queue, skb);
1460 queue->tx.req_cons = idx;
1462 if (((gop-queue->tx_map_ops) >= ARRAY_SIZE(queue->tx_map_ops)) ||
1463 (*copy_ops >= ARRAY_SIZE(queue->tx_copy_ops)))
1467 (*map_ops) = gop - queue->tx_map_ops;
1471 /* Consolidate skb with a frag_list into a brand new one with local pages on
1472 * frags. Returns 0 or -ENOMEM if can't allocate new pages.
1474 static int xenvif_handle_frag_list(struct xenvif_queue *queue, struct sk_buff *skb)
1476 unsigned int offset = skb_headlen(skb);
1477 skb_frag_t frags[MAX_SKB_FRAGS];
1479 struct ubuf_info *uarg;
1480 struct sk_buff *nskb = skb_shinfo(skb)->frag_list;
1482 queue->stats.tx_zerocopy_sent += 2;
1483 queue->stats.tx_frag_overflow++;
1485 xenvif_fill_frags(queue, nskb);
1486 /* Subtract frags size, we will correct it later */
1487 skb->truesize -= skb->data_len;
1488 skb->len += nskb->len;
1489 skb->data_len += nskb->len;
1491 /* create a brand new frags array and coalesce there */
1492 for (i = 0; offset < skb->len; i++) {
1496 BUG_ON(i >= MAX_SKB_FRAGS);
1497 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
1500 skb->truesize += skb->data_len;
1501 for (j = 0; j < i; j++)
1502 put_page(frags[j].page.p);
1506 if (offset + PAGE_SIZE < skb->len)
1509 len = skb->len - offset;
1510 if (skb_copy_bits(skb, offset, page_address(page), len))
1514 frags[i].page.p = page;
1515 frags[i].page_offset = 0;
1516 skb_frag_size_set(&frags[i], len);
1518 /* swap out with old one */
1519 memcpy(skb_shinfo(skb)->frags,
1521 i * sizeof(skb_frag_t));
1522 skb_shinfo(skb)->nr_frags = i;
1523 skb->truesize += i * PAGE_SIZE;
1525 /* remove traces of mapped pages and frag_list */
1526 skb_frag_list_init(skb);
1527 uarg = skb_shinfo(skb)->destructor_arg;
1528 /* increase inflight counter to offset decrement in callback */
1529 atomic_inc(&queue->inflight_packets);
1530 uarg->callback(uarg, true);
1531 skb_shinfo(skb)->destructor_arg = NULL;
1533 xenvif_skb_zerocopy_prepare(queue, nskb);
1539 static int xenvif_tx_submit(struct xenvif_queue *queue)
1541 struct gnttab_map_grant_ref *gop_map = queue->tx_map_ops;
1542 struct gnttab_copy *gop_copy = queue->tx_copy_ops;
1543 struct sk_buff *skb;
1546 while ((skb = __skb_dequeue(&queue->tx_queue)) != NULL) {
1547 struct xen_netif_tx_request *txp;
1551 pending_idx = XENVIF_TX_CB(skb)->pending_idx;
1552 txp = &queue->pending_tx_info[pending_idx].req;
1554 /* Check the remap error code. */
1555 if (unlikely(xenvif_tx_check_gop(queue, skb, &gop_map, &gop_copy))) {
1556 /* If there was an error, xenvif_tx_check_gop is
1557 * expected to release all the frags which were mapped,
1558 * so kfree_skb shouldn't do it again
1560 skb_shinfo(skb)->nr_frags = 0;
1561 if (skb_has_frag_list(skb)) {
1562 struct sk_buff *nskb =
1563 skb_shinfo(skb)->frag_list;
1564 skb_shinfo(nskb)->nr_frags = 0;
1570 data_len = skb->len;
1571 callback_param(queue, pending_idx).ctx = NULL;
1572 if (data_len < txp->size) {
1573 /* Append the packet payload as a fragment. */
1574 txp->offset += data_len;
1575 txp->size -= data_len;
1577 /* Schedule a response immediately. */
1578 xenvif_idx_release(queue, pending_idx,
1579 XEN_NETIF_RSP_OKAY);
1582 if (txp->flags & XEN_NETTXF_csum_blank)
1583 skb->ip_summed = CHECKSUM_PARTIAL;
1584 else if (txp->flags & XEN_NETTXF_data_validated)
1585 skb->ip_summed = CHECKSUM_UNNECESSARY;
1587 xenvif_fill_frags(queue, skb);
1589 if (unlikely(skb_has_frag_list(skb))) {
1590 if (xenvif_handle_frag_list(queue, skb)) {
1591 if (net_ratelimit())
1592 netdev_err(queue->vif->dev,
1593 "Not enough memory to consolidate frag_list!\n");
1594 xenvif_skb_zerocopy_prepare(queue, skb);
1600 if (skb_is_nonlinear(skb) && skb_headlen(skb) < PKT_PROT_LEN) {
1601 int target = min_t(int, skb->len, PKT_PROT_LEN);
1602 __pskb_pull_tail(skb, target - skb_headlen(skb));
1605 skb->dev = queue->vif->dev;
1606 skb->protocol = eth_type_trans(skb, skb->dev);
1607 skb_reset_network_header(skb);
1609 if (checksum_setup(queue, skb)) {
1610 netdev_dbg(queue->vif->dev,
1611 "Can't setup checksum in net_tx_action\n");
1612 /* We have to set this flag to trigger the callback */
1613 if (skb_shinfo(skb)->destructor_arg)
1614 xenvif_skb_zerocopy_prepare(queue, skb);
1619 skb_probe_transport_header(skb, 0);
1621 /* If the packet is GSO then we will have just set up the
1622 * transport header offset in checksum_setup so it's now
1623 * straightforward to calculate gso_segs.
1625 if (skb_is_gso(skb)) {
1626 int mss = skb_shinfo(skb)->gso_size;
1627 int hdrlen = skb_transport_header(skb) -
1628 skb_mac_header(skb) +
1631 skb_shinfo(skb)->gso_segs =
1632 DIV_ROUND_UP(skb->len - hdrlen, mss);
1635 queue->stats.rx_bytes += skb->len;
1636 queue->stats.rx_packets++;
1640 /* Set this flag right before netif_receive_skb, otherwise
1641 * someone might think this packet already left netback, and
1642 * do a skb_copy_ubufs while we are still in control of the
1643 * skb. E.g. the __pskb_pull_tail earlier can do such thing.
1645 if (skb_shinfo(skb)->destructor_arg) {
1646 xenvif_skb_zerocopy_prepare(queue, skb);
1647 queue->stats.tx_zerocopy_sent++;
1650 netif_receive_skb(skb);
1656 void xenvif_zerocopy_callback(struct ubuf_info *ubuf, bool zerocopy_success)
1658 unsigned long flags;
1659 pending_ring_idx_t index;
1660 struct xenvif_queue *queue = ubuf_to_queue(ubuf);
1662 /* This is the only place where we grab this lock, to protect callbacks
1665 spin_lock_irqsave(&queue->callback_lock, flags);
1667 u16 pending_idx = ubuf->desc;
1668 ubuf = (struct ubuf_info *) ubuf->ctx;
1669 BUG_ON(queue->dealloc_prod - queue->dealloc_cons >=
1671 index = pending_index(queue->dealloc_prod);
1672 queue->dealloc_ring[index] = pending_idx;
1673 /* Sync with xenvif_tx_dealloc_action:
1674 * insert idx then incr producer.
1677 queue->dealloc_prod++;
1679 wake_up(&queue->dealloc_wq);
1680 spin_unlock_irqrestore(&queue->callback_lock, flags);
1682 if (likely(zerocopy_success))
1683 queue->stats.tx_zerocopy_success++;
1685 queue->stats.tx_zerocopy_fail++;
1686 xenvif_skb_zerocopy_complete(queue);
1689 static inline void xenvif_tx_dealloc_action(struct xenvif_queue *queue)
1691 struct gnttab_unmap_grant_ref *gop;
1692 pending_ring_idx_t dc, dp;
1693 u16 pending_idx, pending_idx_release[MAX_PENDING_REQS];
1696 dc = queue->dealloc_cons;
1697 gop = queue->tx_unmap_ops;
1699 /* Free up any grants we have finished using */
1701 dp = queue->dealloc_prod;
1703 /* Ensure we see all indices enqueued by all
1704 * xenvif_zerocopy_callback().
1709 BUG_ON(gop - queue->tx_unmap_ops > MAX_PENDING_REQS);
1711 queue->dealloc_ring[pending_index(dc++)];
1713 pending_idx_release[gop-queue->tx_unmap_ops] =
1715 queue->pages_to_unmap[gop-queue->tx_unmap_ops] =
1716 queue->mmap_pages[pending_idx];
1717 gnttab_set_unmap_op(gop,
1718 idx_to_kaddr(queue, pending_idx),
1720 queue->grant_tx_handle[pending_idx]);
1721 xenvif_grant_handle_reset(queue, pending_idx);
1725 } while (dp != queue->dealloc_prod);
1727 queue->dealloc_cons = dc;
1729 if (gop - queue->tx_unmap_ops > 0) {
1731 ret = gnttab_unmap_refs(queue->tx_unmap_ops,
1733 queue->pages_to_unmap,
1734 gop - queue->tx_unmap_ops);
1736 netdev_err(queue->vif->dev, "Unmap fail: nr_ops %tx ret %d\n",
1737 gop - queue->tx_unmap_ops, ret);
1738 for (i = 0; i < gop - queue->tx_unmap_ops; ++i) {
1739 if (gop[i].status != GNTST_okay)
1740 netdev_err(queue->vif->dev,
1741 " host_addr: %llx handle: %x status: %d\n",
1750 for (i = 0; i < gop - queue->tx_unmap_ops; ++i)
1751 xenvif_idx_release(queue, pending_idx_release[i],
1752 XEN_NETIF_RSP_OKAY);
1756 /* Called after netfront has transmitted */
1757 int xenvif_tx_action(struct xenvif_queue *queue, int budget)
1759 unsigned nr_mops, nr_cops = 0;
1762 if (unlikely(!tx_work_todo(queue)))
1765 xenvif_tx_build_gops(queue, budget, &nr_cops, &nr_mops);
1770 gnttab_batch_copy(queue->tx_copy_ops, nr_cops);
1772 ret = gnttab_map_refs(queue->tx_map_ops,
1774 queue->pages_to_map,
1779 work_done = xenvif_tx_submit(queue);
1784 static void xenvif_idx_release(struct xenvif_queue *queue, u16 pending_idx,
1787 struct pending_tx_info *pending_tx_info;
1788 pending_ring_idx_t index;
1789 unsigned long flags;
1791 pending_tx_info = &queue->pending_tx_info[pending_idx];
1792 spin_lock_irqsave(&queue->response_lock, flags);
1793 make_tx_response(queue, &pending_tx_info->req, status);
1794 index = pending_index(queue->pending_prod);
1795 queue->pending_ring[index] = pending_idx;
1796 /* TX shouldn't use the index before we give it back here */
1798 queue->pending_prod++;
1799 spin_unlock_irqrestore(&queue->response_lock, flags);
1803 static void make_tx_response(struct xenvif_queue *queue,
1804 struct xen_netif_tx_request *txp,
1807 RING_IDX i = queue->tx.rsp_prod_pvt;
1808 struct xen_netif_tx_response *resp;
1811 resp = RING_GET_RESPONSE(&queue->tx, i);
1815 if (txp->flags & XEN_NETTXF_extra_info)
1816 RING_GET_RESPONSE(&queue->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1818 queue->tx.rsp_prod_pvt = ++i;
1819 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->tx, notify);
1821 notify_remote_via_irq(queue->tx_irq);
1824 static struct xen_netif_rx_response *make_rx_response(struct xenvif_queue *queue,
1831 RING_IDX i = queue->rx.rsp_prod_pvt;
1832 struct xen_netif_rx_response *resp;
1834 resp = RING_GET_RESPONSE(&queue->rx, i);
1835 resp->offset = offset;
1836 resp->flags = flags;
1838 resp->status = (s16)size;
1840 resp->status = (s16)st;
1842 queue->rx.rsp_prod_pvt = ++i;
1847 void xenvif_idx_unmap(struct xenvif_queue *queue, u16 pending_idx)
1850 struct gnttab_unmap_grant_ref tx_unmap_op;
1852 gnttab_set_unmap_op(&tx_unmap_op,
1853 idx_to_kaddr(queue, pending_idx),
1855 queue->grant_tx_handle[pending_idx]);
1856 xenvif_grant_handle_reset(queue, pending_idx);
1858 ret = gnttab_unmap_refs(&tx_unmap_op, NULL,
1859 &queue->mmap_pages[pending_idx], 1);
1861 netdev_err(queue->vif->dev,
1862 "Unmap fail: ret: %d pending_idx: %d host_addr: %llx handle: %x status: %d\n",
1865 tx_unmap_op.host_addr,
1867 tx_unmap_op.status);
1872 static inline int rx_work_todo(struct xenvif_queue *queue)
1874 return (!skb_queue_empty(&queue->rx_queue) &&
1875 xenvif_rx_ring_slots_available(queue, queue->rx_last_skb_slots));
1878 static inline int tx_work_todo(struct xenvif_queue *queue)
1880 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&queue->tx)))
1886 static inline bool tx_dealloc_work_todo(struct xenvif_queue *queue)
1888 return queue->dealloc_cons != queue->dealloc_prod;
1891 void xenvif_unmap_frontend_rings(struct xenvif_queue *queue)
1893 if (queue->tx.sring)
1894 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1896 if (queue->rx.sring)
1897 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(queue->vif),
1901 int xenvif_map_frontend_rings(struct xenvif_queue *queue,
1902 grant_ref_t tx_ring_ref,
1903 grant_ref_t rx_ring_ref)
1906 struct xen_netif_tx_sring *txs;
1907 struct xen_netif_rx_sring *rxs;
1911 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1912 tx_ring_ref, &addr);
1916 txs = (struct xen_netif_tx_sring *)addr;
1917 BACK_RING_INIT(&queue->tx, txs, PAGE_SIZE);
1919 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(queue->vif),
1920 rx_ring_ref, &addr);
1924 rxs = (struct xen_netif_rx_sring *)addr;
1925 BACK_RING_INIT(&queue->rx, rxs, PAGE_SIZE);
1930 xenvif_unmap_frontend_rings(queue);
1934 static void xenvif_start_queue(struct xenvif_queue *queue)
1936 if (xenvif_schedulable(queue->vif))
1937 xenvif_wake_queue(queue);
1940 /* Only called from the queue's thread, it handles the situation when the guest
1941 * doesn't post enough requests on the receiving ring.
1942 * First xenvif_start_xmit disables QDisc and start a timer, and then either the
1943 * timer fires, or the guest send an interrupt after posting new request. If it
1944 * is the timer, the carrier is turned off here.
1946 static void xenvif_rx_purge_event(struct xenvif_queue *queue)
1948 /* Either the last unsuccesful skb or at least 1 slot should fit */
1949 int needed = queue->rx_last_skb_slots ?
1950 queue->rx_last_skb_slots : 1;
1952 /* It is assumed that if the guest post new slots after this, the RX
1953 * interrupt will set the QUEUE_STATUS_RX_PURGE_EVENT bit and wake up
1956 set_bit(QUEUE_STATUS_RX_STALLED, &queue->status);
1957 if (!xenvif_rx_ring_slots_available(queue, needed)) {
1959 if (netif_carrier_ok(queue->vif->dev)) {
1960 /* Timer fired and there are still no slots. Turn off
1961 * everything except the interrupts
1963 netif_carrier_off(queue->vif->dev);
1964 skb_queue_purge(&queue->rx_queue);
1965 queue->rx_last_skb_slots = 0;
1966 if (net_ratelimit())
1967 netdev_err(queue->vif->dev, "Carrier off due to lack of guest response on queue %d\n", queue->id);
1969 /* Probably an another queue already turned the carrier
1970 * off, make sure nothing is stucked in the internal
1971 * queue of this queue
1973 skb_queue_purge(&queue->rx_queue);
1974 queue->rx_last_skb_slots = 0;
1977 } else if (!netif_carrier_ok(queue->vif->dev)) {
1978 unsigned int num_queues = queue->vif->num_queues;
1980 /* The carrier was down, but an interrupt kicked
1981 * the thread again after new requests were
1984 clear_bit(QUEUE_STATUS_RX_STALLED,
1987 netif_carrier_on(queue->vif->dev);
1988 netif_tx_wake_all_queues(queue->vif->dev);
1991 for (i = 0; i < num_queues; i++) {
1992 struct xenvif_queue *temp = &queue->vif->queues[i];
1994 xenvif_napi_schedule_or_enable_events(temp);
1996 if (net_ratelimit())
1997 netdev_err(queue->vif->dev, "Carrier on again\n");
1999 /* Queuing were stopped, but the guest posted
2000 * new requests and sent an interrupt
2002 clear_bit(QUEUE_STATUS_RX_STALLED,
2004 del_timer_sync(&queue->rx_stalled);
2005 xenvif_start_queue(queue);
2009 int xenvif_kthread_guest_rx(void *data)
2011 struct xenvif_queue *queue = data;
2012 struct sk_buff *skb;
2014 while (!kthread_should_stop()) {
2015 wait_event_interruptible(queue->wq,
2016 rx_work_todo(queue) ||
2017 queue->vif->disabled ||
2018 test_bit(QUEUE_STATUS_RX_PURGE_EVENT, &queue->status) ||
2019 kthread_should_stop());
2021 if (kthread_should_stop())
2024 /* This frontend is found to be rogue, disable it in
2025 * kthread context. Currently this is only set when
2026 * netback finds out frontend sends malformed packet,
2027 * but we cannot disable the interface in softirq
2028 * context so we defer it here, if this thread is
2029 * associated with queue 0.
2031 if (unlikely(queue->vif->disabled && queue->id == 0)) {
2032 xenvif_carrier_off(queue->vif);
2033 } else if (unlikely(queue->vif->disabled)) {
2034 /* kthread_stop() would be called upon this thread soon,
2035 * be a bit proactive
2037 skb_queue_purge(&queue->rx_queue);
2038 queue->rx_last_skb_slots = 0;
2039 } else if (unlikely(test_and_clear_bit(QUEUE_STATUS_RX_PURGE_EVENT,
2041 xenvif_rx_purge_event(queue);
2042 } else if (!netif_carrier_ok(queue->vif->dev)) {
2043 /* Another queue stalled and turned the carrier off, so
2044 * purge the internal queue of queues which were not
2047 skb_queue_purge(&queue->rx_queue);
2048 queue->rx_last_skb_slots = 0;
2051 if (!skb_queue_empty(&queue->rx_queue))
2052 xenvif_rx_action(queue);
2057 /* Bin any remaining skbs */
2058 while ((skb = skb_dequeue(&queue->rx_queue)) != NULL)
2064 static bool xenvif_dealloc_kthread_should_stop(struct xenvif_queue *queue)
2066 /* Dealloc thread must remain running until all inflight
2069 return kthread_should_stop() &&
2070 !atomic_read(&queue->inflight_packets);
2073 int xenvif_dealloc_kthread(void *data)
2075 struct xenvif_queue *queue = data;
2078 wait_event_interruptible(queue->dealloc_wq,
2079 tx_dealloc_work_todo(queue) ||
2080 xenvif_dealloc_kthread_should_stop(queue));
2081 if (xenvif_dealloc_kthread_should_stop(queue))
2084 xenvif_tx_dealloc_action(queue);
2088 /* Unmap anything remaining*/
2089 if (tx_dealloc_work_todo(queue))
2090 xenvif_tx_dealloc_action(queue);
2095 static int __init netback_init(void)
2102 /* Allow as many queues as there are CPUs, by default */
2103 xenvif_max_queues = num_online_cpus();
2105 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
2106 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
2107 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
2108 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
2111 rc = xenvif_xenbus_init();
2115 rx_drain_timeout_jiffies = msecs_to_jiffies(rx_drain_timeout_msecs);
2117 #ifdef CONFIG_DEBUG_FS
2118 xen_netback_dbg_root = debugfs_create_dir("xen-netback", NULL);
2119 if (IS_ERR_OR_NULL(xen_netback_dbg_root))
2120 pr_warn("Init of debugfs returned %ld!\n",
2121 PTR_ERR(xen_netback_dbg_root));
2122 #endif /* CONFIG_DEBUG_FS */
2130 module_init(netback_init);
2132 static void __exit netback_fini(void)
2134 #ifdef CONFIG_DEBUG_FS
2135 if (!IS_ERR_OR_NULL(xen_netback_dbg_root))
2136 debugfs_remove_recursive(xen_netback_dbg_root);
2137 #endif /* CONFIG_DEBUG_FS */
2138 xenvif_xenbus_fini();
2140 module_exit(netback_fini);
2142 MODULE_LICENSE("Dual BSD/GPL");
2143 MODULE_ALIAS("xen-backend:vif");