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>
44 #include <xen/events.h>
45 #include <xen/interface/memory.h>
47 #include <asm/xen/hypercall.h>
48 #include <asm/xen/page.h>
50 /* Provide an option to disable split event channels at load time as
51 * event channels are limited resource. Split event channels are
54 bool separate_tx_rx_irq = 1;
55 module_param(separate_tx_rx_irq, bool, 0644);
58 * This is the maximum slots a skb can have. If a guest sends a skb
59 * which exceeds this limit it is considered malicious.
61 #define FATAL_SKB_SLOTS_DEFAULT 20
62 static unsigned int fatal_skb_slots = FATAL_SKB_SLOTS_DEFAULT;
63 module_param(fatal_skb_slots, uint, 0444);
66 * To avoid confusion, we define XEN_NETBK_LEGACY_SLOTS_MAX indicating
67 * the maximum slots a valid packet can use. Now this value is defined
68 * to be XEN_NETIF_NR_SLOTS_MIN, which is supposed to be supported by
71 #define XEN_NETBK_LEGACY_SLOTS_MAX XEN_NETIF_NR_SLOTS_MIN
74 * If head != INVALID_PENDING_RING_IDX, it means this tx request is head of
75 * one or more merged tx requests, otherwise it is the continuation of
76 * previous tx request.
78 static inline int pending_tx_is_head(struct xenvif *vif, RING_IDX idx)
80 return vif->pending_tx_info[idx].head != INVALID_PENDING_RING_IDX;
83 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
86 static void make_tx_response(struct xenvif *vif,
87 struct xen_netif_tx_request *txp,
90 static inline int tx_work_todo(struct xenvif *vif);
91 static inline int rx_work_todo(struct xenvif *vif);
93 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
100 static inline unsigned long idx_to_pfn(struct xenvif *vif,
103 return page_to_pfn(vif->mmap_pages[idx]);
106 static inline unsigned long idx_to_kaddr(struct xenvif *vif,
109 return (unsigned long)pfn_to_kaddr(idx_to_pfn(vif, idx));
113 * This is the amount of packet we copy rather than map, so that the
114 * guest can't fiddle with the contents of the headers while we do
115 * packet processing on them (netfilter, routing, etc).
117 #define PKT_PROT_LEN (ETH_HLEN + \
119 sizeof(struct iphdr) + MAX_IPOPTLEN + \
120 sizeof(struct tcphdr) + MAX_TCP_OPTION_SPACE)
122 static u16 frag_get_pending_idx(skb_frag_t *frag)
124 return (u16)frag->page_offset;
127 static void frag_set_pending_idx(skb_frag_t *frag, u16 pending_idx)
129 frag->page_offset = pending_idx;
132 static inline pending_ring_idx_t pending_index(unsigned i)
134 return i & (MAX_PENDING_REQS-1);
137 static inline pending_ring_idx_t nr_pending_reqs(struct xenvif *vif)
139 return MAX_PENDING_REQS -
140 vif->pending_prod + vif->pending_cons;
143 static int max_required_rx_slots(struct xenvif *vif)
145 int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
147 /* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
148 if (vif->can_sg || vif->gso || vif->gso_prefix)
149 max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
154 int xenvif_rx_ring_full(struct xenvif *vif)
156 RING_IDX peek = vif->rx_req_cons_peek;
157 RING_IDX needed = max_required_rx_slots(vif);
159 return ((vif->rx.sring->req_prod - peek) < needed) ||
160 ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
163 int xenvif_must_stop_queue(struct xenvif *vif)
165 if (!xenvif_rx_ring_full(vif))
168 vif->rx.sring->req_event = vif->rx_req_cons_peek +
169 max_required_rx_slots(vif);
170 mb(); /* request notification /then/ check the queue */
172 return xenvif_rx_ring_full(vif);
176 * Returns true if we should start a new receive buffer instead of
177 * adding 'size' bytes to a buffer which currently contains 'offset'
180 static bool start_new_rx_buffer(int offset, unsigned long size, int head)
182 /* simple case: we have completely filled the current buffer. */
183 if (offset == MAX_BUFFER_OFFSET)
187 * complex case: start a fresh buffer if the current frag
188 * would overflow the current buffer but only if:
189 * (i) this frag would fit completely in the next buffer
190 * and (ii) there is already some data in the current buffer
191 * and (iii) this is not the head buffer.
194 * - (i) stops us splitting a frag into two copies
195 * unless the frag is too large for a single buffer.
196 * - (ii) stops us from leaving a buffer pointlessly empty.
197 * - (iii) stops us leaving the first buffer
198 * empty. Strictly speaking this is already covered
199 * by (ii) but is explicitly checked because
200 * netfront relies on the first buffer being
201 * non-empty and can crash otherwise.
203 * This means we will effectively linearise small
204 * frags but do not needlessly split large buffers
205 * into multiple copies tend to give large frags their
206 * own buffers as before.
208 if ((offset + size > MAX_BUFFER_OFFSET) &&
209 (size <= MAX_BUFFER_OFFSET) && offset && !head)
216 * Figure out how many ring slots we're going to need to send @skb to
217 * the guest. This function is essentially a dry run of
218 * xenvif_gop_frag_copy.
220 unsigned int xenvif_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
225 count = DIV_ROUND_UP(skb_headlen(skb), PAGE_SIZE);
227 copy_off = skb_headlen(skb) % PAGE_SIZE;
229 if (skb_shinfo(skb)->gso_size)
232 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
233 unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
234 unsigned long offset = skb_shinfo(skb)->frags[i].page_offset;
237 offset &= ~PAGE_MASK;
240 BUG_ON(offset >= PAGE_SIZE);
241 BUG_ON(copy_off > MAX_BUFFER_OFFSET);
243 bytes = PAGE_SIZE - offset;
248 if (start_new_rx_buffer(copy_off, bytes, 0)) {
253 if (copy_off + bytes > MAX_BUFFER_OFFSET)
254 bytes = MAX_BUFFER_OFFSET - copy_off;
261 if (offset == PAGE_SIZE)
268 struct netrx_pending_operations {
269 unsigned copy_prod, copy_cons;
270 unsigned meta_prod, meta_cons;
271 struct gnttab_copy *copy;
272 struct xenvif_rx_meta *meta;
274 grant_ref_t copy_gref;
277 static struct xenvif_rx_meta *get_next_rx_buffer(struct xenvif *vif,
278 struct netrx_pending_operations *npo)
280 struct xenvif_rx_meta *meta;
281 struct xen_netif_rx_request *req;
283 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
285 meta = npo->meta + npo->meta_prod++;
291 npo->copy_gref = req->gref;
297 * Set up the grant operations for this fragment. If it's a flipping
298 * interface, we also set up the unmap request from here.
300 static void xenvif_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
301 struct netrx_pending_operations *npo,
302 struct page *page, unsigned long size,
303 unsigned long offset, int *head)
305 struct gnttab_copy *copy_gop;
306 struct xenvif_rx_meta *meta;
309 /* Data must not cross a page boundary. */
310 BUG_ON(size + offset > PAGE_SIZE<<compound_order(page));
312 meta = npo->meta + npo->meta_prod - 1;
314 /* Skip unused frames from start of page */
315 page += offset >> PAGE_SHIFT;
316 offset &= ~PAGE_MASK;
319 BUG_ON(offset >= PAGE_SIZE);
320 BUG_ON(npo->copy_off > MAX_BUFFER_OFFSET);
322 bytes = PAGE_SIZE - offset;
327 if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
329 * Netfront requires there to be some data in the head
334 meta = get_next_rx_buffer(vif, npo);
337 if (npo->copy_off + bytes > MAX_BUFFER_OFFSET)
338 bytes = MAX_BUFFER_OFFSET - npo->copy_off;
340 copy_gop = npo->copy + npo->copy_prod++;
341 copy_gop->flags = GNTCOPY_dest_gref;
342 copy_gop->len = bytes;
344 copy_gop->source.domid = DOMID_SELF;
345 copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
346 copy_gop->source.offset = offset;
348 copy_gop->dest.domid = vif->domid;
349 copy_gop->dest.offset = npo->copy_off;
350 copy_gop->dest.u.ref = npo->copy_gref;
352 npo->copy_off += bytes;
359 if (offset == PAGE_SIZE && size) {
360 BUG_ON(!PageCompound(page));
365 /* Leave a gap for the GSO descriptor. */
366 if (*head && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
369 *head = 0; /* There must be something in this buffer now. */
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)
389 struct xenvif *vif = netdev_priv(skb->dev);
390 int nr_frags = skb_shinfo(skb)->nr_frags;
392 struct xen_netif_rx_request *req;
393 struct xenvif_rx_meta *meta;
398 old_meta_prod = npo->meta_prod;
400 /* Set up a GSO prefix descriptor, if necessary */
401 if (skb_shinfo(skb)->gso_size && vif->gso_prefix) {
402 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
403 meta = npo->meta + npo->meta_prod++;
404 meta->gso_size = skb_shinfo(skb)->gso_size;
409 req = RING_GET_REQUEST(&vif->rx, vif->rx.req_cons++);
410 meta = npo->meta + npo->meta_prod++;
412 if (!vif->gso_prefix)
413 meta->gso_size = skb_shinfo(skb)->gso_size;
420 npo->copy_gref = req->gref;
423 while (data < skb_tail_pointer(skb)) {
424 unsigned int offset = offset_in_page(data);
425 unsigned int len = PAGE_SIZE - offset;
427 if (data + len > skb_tail_pointer(skb))
428 len = skb_tail_pointer(skb) - data;
430 xenvif_gop_frag_copy(vif, skb, npo,
431 virt_to_page(data), len, offset, &head);
435 for (i = 0; i < nr_frags; i++) {
436 xenvif_gop_frag_copy(vif, skb, npo,
437 skb_frag_page(&skb_shinfo(skb)->frags[i]),
438 skb_frag_size(&skb_shinfo(skb)->frags[i]),
439 skb_shinfo(skb)->frags[i].page_offset,
443 return npo->meta_prod - old_meta_prod;
447 * This is a twin to xenvif_gop_skb. Assume that xenvif_gop_skb was
448 * used to set up the operations on the top of
449 * netrx_pending_operations, which have since been done. Check that
450 * they didn't give any errors and advance over them.
452 static int xenvif_check_gop(struct xenvif *vif, int nr_meta_slots,
453 struct netrx_pending_operations *npo)
455 struct gnttab_copy *copy_op;
456 int status = XEN_NETIF_RSP_OKAY;
459 for (i = 0; i < nr_meta_slots; i++) {
460 copy_op = npo->copy + npo->copy_cons++;
461 if (copy_op->status != GNTST_okay) {
463 "Bad status %d from copy to DOM%d.\n",
464 copy_op->status, vif->domid);
465 status = XEN_NETIF_RSP_ERROR;
472 static void xenvif_add_frag_responses(struct xenvif *vif, int status,
473 struct xenvif_rx_meta *meta,
477 unsigned long offset;
479 /* No fragments used */
480 if (nr_meta_slots <= 1)
485 for (i = 0; i < nr_meta_slots; i++) {
487 if (i == nr_meta_slots - 1)
490 flags = XEN_NETRXF_more_data;
493 make_rx_response(vif, meta[i].id, status, offset,
494 meta[i].size, flags);
498 struct skb_cb_overlay {
502 static void xenvif_kick_thread(struct xenvif *vif)
507 void xenvif_rx_action(struct xenvif *vif)
511 struct xen_netif_rx_response *resp;
512 struct sk_buff_head rxq;
518 unsigned long offset;
519 struct skb_cb_overlay *sco;
520 int need_to_notify = 0;
522 struct netrx_pending_operations npo = {
523 .copy = vif->grant_copy_op,
527 skb_queue_head_init(&rxq);
531 while ((skb = skb_dequeue(&vif->rx_queue)) != NULL) {
532 vif = netdev_priv(skb->dev);
533 nr_frags = skb_shinfo(skb)->nr_frags;
535 sco = (struct skb_cb_overlay *)skb->cb;
536 sco->meta_slots_used = xenvif_gop_skb(skb, &npo);
538 count += nr_frags + 1;
540 __skb_queue_tail(&rxq, skb);
542 /* Filled the batch queue? */
543 /* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
544 if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
548 BUG_ON(npo.meta_prod > ARRAY_SIZE(vif->meta));
553 BUG_ON(npo.copy_prod > ARRAY_SIZE(vif->grant_copy_op));
554 gnttab_batch_copy(vif->grant_copy_op, npo.copy_prod);
556 while ((skb = __skb_dequeue(&rxq)) != NULL) {
557 sco = (struct skb_cb_overlay *)skb->cb;
559 vif = netdev_priv(skb->dev);
561 if (vif->meta[npo.meta_cons].gso_size && vif->gso_prefix) {
562 resp = RING_GET_RESPONSE(&vif->rx,
563 vif->rx.rsp_prod_pvt++);
565 resp->flags = XEN_NETRXF_gso_prefix | XEN_NETRXF_more_data;
567 resp->offset = vif->meta[npo.meta_cons].gso_size;
568 resp->id = vif->meta[npo.meta_cons].id;
569 resp->status = sco->meta_slots_used;
572 sco->meta_slots_used--;
576 vif->dev->stats.tx_bytes += skb->len;
577 vif->dev->stats.tx_packets++;
579 status = xenvif_check_gop(vif, sco->meta_slots_used, &npo);
581 if (sco->meta_slots_used == 1)
584 flags = XEN_NETRXF_more_data;
586 if (skb->ip_summed == CHECKSUM_PARTIAL) /* local packet? */
587 flags |= XEN_NETRXF_csum_blank | XEN_NETRXF_data_validated;
588 else if (skb->ip_summed == CHECKSUM_UNNECESSARY)
589 /* remote but checksummed. */
590 flags |= XEN_NETRXF_data_validated;
593 resp = make_rx_response(vif, vif->meta[npo.meta_cons].id,
595 vif->meta[npo.meta_cons].size,
598 if (vif->meta[npo.meta_cons].gso_size && !vif->gso_prefix) {
599 struct xen_netif_extra_info *gso =
600 (struct xen_netif_extra_info *)
601 RING_GET_RESPONSE(&vif->rx,
602 vif->rx.rsp_prod_pvt++);
604 resp->flags |= XEN_NETRXF_extra_info;
606 gso->u.gso.size = vif->meta[npo.meta_cons].gso_size;
607 gso->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
609 gso->u.gso.features = 0;
611 gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
615 xenvif_add_frag_responses(vif, status,
616 vif->meta + npo.meta_cons + 1,
617 sco->meta_slots_used);
619 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->rx, ret);
624 xenvif_notify_tx_completion(vif);
626 npo.meta_cons += sco->meta_slots_used;
631 notify_remote_via_irq(vif->rx_irq);
633 /* More work to do? */
634 if (!skb_queue_empty(&vif->rx_queue))
635 xenvif_kick_thread(vif);
638 void xenvif_queue_tx_skb(struct xenvif *vif, struct sk_buff *skb)
640 skb_queue_tail(&vif->rx_queue, skb);
642 xenvif_kick_thread(vif);
645 void xenvif_check_rx_xenvif(struct xenvif *vif)
649 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
652 napi_schedule(&vif->napi);
655 static void tx_add_credit(struct xenvif *vif)
657 unsigned long max_burst, max_credit;
660 * Allow a burst big enough to transmit a jumbo packet of up to 128kB.
661 * Otherwise the interface can seize up due to insufficient credit.
663 max_burst = RING_GET_REQUEST(&vif->tx, vif->tx.req_cons)->size;
664 max_burst = min(max_burst, 131072UL);
665 max_burst = max(max_burst, vif->credit_bytes);
667 /* Take care that adding a new chunk of credit doesn't wrap to zero. */
668 max_credit = vif->remaining_credit + vif->credit_bytes;
669 if (max_credit < vif->remaining_credit)
670 max_credit = ULONG_MAX; /* wrapped: clamp to ULONG_MAX */
672 vif->remaining_credit = min(max_credit, max_burst);
675 static void tx_credit_callback(unsigned long data)
677 struct xenvif *vif = (struct xenvif *)data;
679 xenvif_check_rx_xenvif(vif);
682 static void xenvif_tx_err(struct xenvif *vif,
683 struct xen_netif_tx_request *txp, RING_IDX end)
685 RING_IDX cons = vif->tx.req_cons;
688 make_tx_response(vif, txp, XEN_NETIF_RSP_ERROR);
691 txp = RING_GET_REQUEST(&vif->tx, cons++);
693 vif->tx.req_cons = cons;
696 static void xenvif_fatal_tx_err(struct xenvif *vif)
698 netdev_err(vif->dev, "fatal error; disabling device\n");
699 xenvif_carrier_off(vif);
702 static int xenvif_count_requests(struct xenvif *vif,
703 struct xen_netif_tx_request *first,
704 struct xen_netif_tx_request *txp,
707 RING_IDX cons = vif->tx.req_cons;
712 if (!(first->flags & XEN_NETTXF_more_data))
716 struct xen_netif_tx_request dropped_tx = { 0 };
718 if (slots >= work_to_do) {
720 "Asked for %d slots but exceeds this limit\n",
722 xenvif_fatal_tx_err(vif);
726 /* This guest is really using too many slots and
727 * considered malicious.
729 if (unlikely(slots >= fatal_skb_slots)) {
731 "Malicious frontend using %d slots, threshold %u\n",
732 slots, fatal_skb_slots);
733 xenvif_fatal_tx_err(vif);
737 /* Xen network protocol had implicit dependency on
738 * MAX_SKB_FRAGS. XEN_NETBK_LEGACY_SLOTS_MAX is set to
739 * the historical MAX_SKB_FRAGS value 18 to honor the
740 * same behavior as before. Any packet using more than
741 * 18 slots but less than fatal_skb_slots slots is
744 if (!drop_err && slots >= XEN_NETBK_LEGACY_SLOTS_MAX) {
747 "Too many slots (%d) exceeding limit (%d), dropping packet\n",
748 slots, XEN_NETBK_LEGACY_SLOTS_MAX);
755 memcpy(txp, RING_GET_REQUEST(&vif->tx, cons + slots),
758 /* If the guest submitted a frame >= 64 KiB then
759 * first->size overflowed and following slots will
760 * appear to be larger than the frame.
762 * This cannot be fatal error as there are buggy
763 * frontends that do this.
765 * Consume all slots and drop the packet.
767 if (!drop_err && txp->size > first->size) {
770 "Invalid tx request, slot size %u > remaining size %u\n",
771 txp->size, first->size);
775 first->size -= txp->size;
778 if (unlikely((txp->offset + txp->size) > PAGE_SIZE)) {
779 netdev_err(vif->dev, "Cross page boundary, txp->offset: %x, size: %u\n",
780 txp->offset, txp->size);
781 xenvif_fatal_tx_err(vif);
785 more_data = txp->flags & XEN_NETTXF_more_data;
793 xenvif_tx_err(vif, first, cons + slots);
800 static struct page *xenvif_alloc_page(struct xenvif *vif,
805 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
808 vif->mmap_pages[pending_idx] = page;
813 static struct gnttab_copy *xenvif_get_requests(struct xenvif *vif,
815 struct xen_netif_tx_request *txp,
816 struct gnttab_copy *gop)
818 struct skb_shared_info *shinfo = skb_shinfo(skb);
819 skb_frag_t *frags = shinfo->frags;
820 u16 pending_idx = *((u16 *)skb->data);
824 pending_ring_idx_t index, start_idx = 0;
826 unsigned int nr_slots;
827 struct pending_tx_info *first = NULL;
829 /* At this point shinfo->nr_frags is in fact the number of
830 * slots, which can be as large as XEN_NETBK_LEGACY_SLOTS_MAX.
832 nr_slots = shinfo->nr_frags;
834 /* Skip first skb fragment if it is on same page as header fragment. */
835 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
837 /* Coalesce tx requests, at this point the packet passed in
838 * should be <= 64K. Any packets larger than 64K have been
839 * handled in xenvif_count_requests().
841 for (shinfo->nr_frags = slot = start; slot < nr_slots;
842 shinfo->nr_frags++) {
843 struct pending_tx_info *pending_tx_info =
844 vif->pending_tx_info;
846 page = alloc_page(GFP_ATOMIC|__GFP_COLD);
852 while (dst_offset < PAGE_SIZE && slot < nr_slots) {
853 gop->flags = GNTCOPY_source_gref;
855 gop->source.u.ref = txp->gref;
856 gop->source.domid = vif->domid;
857 gop->source.offset = txp->offset;
859 gop->dest.domid = DOMID_SELF;
861 gop->dest.offset = dst_offset;
862 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
864 if (dst_offset + txp->size > PAGE_SIZE) {
865 /* This page can only merge a portion
866 * of tx request. Do not increment any
867 * pointer / counter here. The txp
868 * will be dealt with in future
869 * rounds, eventually hitting the
872 gop->len = PAGE_SIZE - dst_offset;
873 txp->offset += gop->len;
874 txp->size -= gop->len;
875 dst_offset += gop->len; /* quit loop */
877 /* This tx request can be merged in the page */
878 gop->len = txp->size;
879 dst_offset += gop->len;
881 index = pending_index(vif->pending_cons++);
883 pending_idx = vif->pending_ring[index];
885 memcpy(&pending_tx_info[pending_idx].req, txp,
888 /* Poison these fields, corresponding
889 * fields for head tx req will be set
890 * to correct values after the loop.
892 vif->mmap_pages[pending_idx] = (void *)(~0UL);
893 pending_tx_info[pending_idx].head =
894 INVALID_PENDING_RING_IDX;
897 first = &pending_tx_info[pending_idx];
899 head_idx = pending_idx;
909 first->req.offset = 0;
910 first->req.size = dst_offset;
911 first->head = start_idx;
912 vif->mmap_pages[head_idx] = page;
913 frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
916 BUG_ON(shinfo->nr_frags > MAX_SKB_FRAGS);
920 /* Unwind, freeing all pages and sending error responses. */
921 while (shinfo->nr_frags-- > start) {
922 xenvif_idx_release(vif,
923 frag_get_pending_idx(&frags[shinfo->nr_frags]),
924 XEN_NETIF_RSP_ERROR);
926 /* The head too, if necessary. */
928 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
933 static int xenvif_tx_check_gop(struct xenvif *vif,
935 struct gnttab_copy **gopp)
937 struct gnttab_copy *gop = *gopp;
938 u16 pending_idx = *((u16 *)skb->data);
939 struct skb_shared_info *shinfo = skb_shinfo(skb);
940 struct pending_tx_info *tx_info;
941 int nr_frags = shinfo->nr_frags;
943 u16 peek; /* peek into next tx request */
945 /* Check status of header. */
948 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
950 /* Skip first skb fragment if it is on same page as header fragment. */
951 start = (frag_get_pending_idx(&shinfo->frags[0]) == pending_idx);
953 for (i = start; i < nr_frags; i++) {
955 pending_ring_idx_t head;
957 pending_idx = frag_get_pending_idx(&shinfo->frags[i]);
958 tx_info = &vif->pending_tx_info[pending_idx];
959 head = tx_info->head;
961 /* Check error status: if okay then remember grant handle. */
963 newerr = (++gop)->status;
966 peek = vif->pending_ring[pending_index(++head)];
967 } while (!pending_tx_is_head(vif, peek));
969 if (likely(!newerr)) {
970 /* Had a previous error? Invalidate this fragment. */
972 xenvif_idx_release(vif, pending_idx,
977 /* Error on this fragment: respond to client with an error. */
978 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_ERROR);
980 /* Not the first error? Preceding frags already invalidated. */
984 /* First error: invalidate header and preceding fragments. */
985 pending_idx = *((u16 *)skb->data);
986 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
987 for (j = start; j < i; j++) {
988 pending_idx = frag_get_pending_idx(&shinfo->frags[j]);
989 xenvif_idx_release(vif, pending_idx,
993 /* Remember the error: invalidate all subsequent fragments. */
1001 static void xenvif_fill_frags(struct xenvif *vif, struct sk_buff *skb)
1003 struct skb_shared_info *shinfo = skb_shinfo(skb);
1004 int nr_frags = shinfo->nr_frags;
1007 for (i = 0; i < nr_frags; i++) {
1008 skb_frag_t *frag = shinfo->frags + i;
1009 struct xen_netif_tx_request *txp;
1013 pending_idx = frag_get_pending_idx(frag);
1015 txp = &vif->pending_tx_info[pending_idx].req;
1016 page = virt_to_page(idx_to_kaddr(vif, pending_idx));
1017 __skb_fill_page_desc(skb, i, page, txp->offset, txp->size);
1018 skb->len += txp->size;
1019 skb->data_len += txp->size;
1020 skb->truesize += txp->size;
1022 /* Take an extra reference to offset xenvif_idx_release */
1023 get_page(vif->mmap_pages[pending_idx]);
1024 xenvif_idx_release(vif, pending_idx, XEN_NETIF_RSP_OKAY);
1028 static int xenvif_get_extras(struct xenvif *vif,
1029 struct xen_netif_extra_info *extras,
1032 struct xen_netif_extra_info extra;
1033 RING_IDX cons = vif->tx.req_cons;
1036 if (unlikely(work_to_do-- <= 0)) {
1037 netdev_err(vif->dev, "Missing extra info\n");
1038 xenvif_fatal_tx_err(vif);
1042 memcpy(&extra, RING_GET_REQUEST(&vif->tx, cons),
1044 if (unlikely(!extra.type ||
1045 extra.type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1046 vif->tx.req_cons = ++cons;
1047 netdev_err(vif->dev,
1048 "Invalid extra type: %d\n", extra.type);
1049 xenvif_fatal_tx_err(vif);
1053 memcpy(&extras[extra.type - 1], &extra, sizeof(extra));
1054 vif->tx.req_cons = ++cons;
1055 } while (extra.flags & XEN_NETIF_EXTRA_FLAG_MORE);
1060 static int xenvif_set_skb_gso(struct xenvif *vif,
1061 struct sk_buff *skb,
1062 struct xen_netif_extra_info *gso)
1064 if (!gso->u.gso.size) {
1065 netdev_err(vif->dev, "GSO size must not be zero.\n");
1066 xenvif_fatal_tx_err(vif);
1070 /* Currently only TCPv4 S.O. is supported. */
1071 if (gso->u.gso.type != XEN_NETIF_GSO_TYPE_TCPV4) {
1072 netdev_err(vif->dev, "Bad GSO type %d.\n", gso->u.gso.type);
1073 xenvif_fatal_tx_err(vif);
1077 skb_shinfo(skb)->gso_size = gso->u.gso.size;
1078 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
1080 /* Header must be checked, and gso_segs computed. */
1081 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
1082 skb_shinfo(skb)->gso_segs = 0;
1087 static int checksum_setup(struct xenvif *vif, struct sk_buff *skb)
1091 int recalculate_partial_csum = 0;
1094 * A GSO SKB must be CHECKSUM_PARTIAL. However some buggy
1095 * peers can fail to set NETRXF_csum_blank when sending a GSO
1096 * frame. In this case force the SKB to CHECKSUM_PARTIAL and
1097 * recalculate the partial checksum.
1099 if (skb->ip_summed != CHECKSUM_PARTIAL && skb_is_gso(skb)) {
1100 vif->rx_gso_checksum_fixup++;
1101 skb->ip_summed = CHECKSUM_PARTIAL;
1102 recalculate_partial_csum = 1;
1105 /* A non-CHECKSUM_PARTIAL SKB does not require setup. */
1106 if (skb->ip_summed != CHECKSUM_PARTIAL)
1109 if (skb->protocol != htons(ETH_P_IP))
1112 iph = (void *)skb->data;
1113 switch (iph->protocol) {
1115 if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1116 offsetof(struct tcphdr, check)))
1119 if (recalculate_partial_csum) {
1120 struct tcphdr *tcph = tcp_hdr(skb);
1121 tcph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1122 skb->len - iph->ihl*4,
1127 if (!skb_partial_csum_set(skb, 4 * iph->ihl,
1128 offsetof(struct udphdr, check)))
1131 if (recalculate_partial_csum) {
1132 struct udphdr *udph = udp_hdr(skb);
1133 udph->check = ~csum_tcpudp_magic(iph->saddr, iph->daddr,
1134 skb->len - iph->ihl*4,
1139 if (net_ratelimit())
1140 netdev_err(vif->dev,
1141 "Attempting to checksum a non-TCP/UDP packet, dropping a protocol %d packet\n",
1152 static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
1154 unsigned long now = jiffies;
1155 unsigned long next_credit =
1156 vif->credit_timeout.expires +
1157 msecs_to_jiffies(vif->credit_usec / 1000);
1159 /* Timer could already be pending in rare cases. */
1160 if (timer_pending(&vif->credit_timeout))
1163 /* Passed the point where we can replenish credit? */
1164 if (time_after_eq(now, next_credit)) {
1165 vif->credit_timeout.expires = now;
1169 /* Still too big to send right now? Set a callback. */
1170 if (size > vif->remaining_credit) {
1171 vif->credit_timeout.data =
1173 vif->credit_timeout.function =
1175 mod_timer(&vif->credit_timeout,
1184 static unsigned xenvif_tx_build_gops(struct xenvif *vif)
1186 struct gnttab_copy *gop = vif->tx_copy_ops, *request_gop;
1187 struct sk_buff *skb;
1190 while ((nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
1191 < MAX_PENDING_REQS)) {
1192 struct xen_netif_tx_request txreq;
1193 struct xen_netif_tx_request txfrags[XEN_NETBK_LEGACY_SLOTS_MAX];
1195 struct xen_netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX-1];
1199 unsigned int data_len;
1200 pending_ring_idx_t index;
1202 if (vif->tx.sring->req_prod - vif->tx.req_cons >
1203 XEN_NETIF_TX_RING_SIZE) {
1204 netdev_err(vif->dev,
1205 "Impossible number of requests. "
1206 "req_prod %d, req_cons %d, size %ld\n",
1207 vif->tx.sring->req_prod, vif->tx.req_cons,
1208 XEN_NETIF_TX_RING_SIZE);
1209 xenvif_fatal_tx_err(vif);
1213 RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, work_to_do);
1217 idx = vif->tx.req_cons;
1218 rmb(); /* Ensure that we see the request before we copy it. */
1219 memcpy(&txreq, RING_GET_REQUEST(&vif->tx, idx), sizeof(txreq));
1221 /* Credit-based scheduling. */
1222 if (txreq.size > vif->remaining_credit &&
1223 tx_credit_exceeded(vif, txreq.size))
1226 vif->remaining_credit -= txreq.size;
1229 vif->tx.req_cons = ++idx;
1231 memset(extras, 0, sizeof(extras));
1232 if (txreq.flags & XEN_NETTXF_extra_info) {
1233 work_to_do = xenvif_get_extras(vif, extras,
1235 idx = vif->tx.req_cons;
1236 if (unlikely(work_to_do < 0))
1240 ret = xenvif_count_requests(vif, &txreq, txfrags, work_to_do);
1241 if (unlikely(ret < 0))
1246 if (unlikely(txreq.size < ETH_HLEN)) {
1247 netdev_dbg(vif->dev,
1248 "Bad packet size: %d\n", txreq.size);
1249 xenvif_tx_err(vif, &txreq, idx);
1253 /* No crossing a page as the payload mustn't fragment. */
1254 if (unlikely((txreq.offset + txreq.size) > PAGE_SIZE)) {
1255 netdev_err(vif->dev,
1256 "txreq.offset: %x, size: %u, end: %lu\n",
1257 txreq.offset, txreq.size,
1258 (txreq.offset&~PAGE_MASK) + txreq.size);
1259 xenvif_fatal_tx_err(vif);
1263 index = pending_index(vif->pending_cons);
1264 pending_idx = vif->pending_ring[index];
1266 data_len = (txreq.size > PKT_PROT_LEN &&
1267 ret < XEN_NETBK_LEGACY_SLOTS_MAX) ?
1268 PKT_PROT_LEN : txreq.size;
1270 skb = alloc_skb(data_len + NET_SKB_PAD + NET_IP_ALIGN,
1271 GFP_ATOMIC | __GFP_NOWARN);
1272 if (unlikely(skb == NULL)) {
1273 netdev_dbg(vif->dev,
1274 "Can't allocate a skb in start_xmit.\n");
1275 xenvif_tx_err(vif, &txreq, idx);
1279 /* Packets passed to netif_rx() must have some headroom. */
1280 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN);
1282 if (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type) {
1283 struct xen_netif_extra_info *gso;
1284 gso = &extras[XEN_NETIF_EXTRA_TYPE_GSO - 1];
1286 if (xenvif_set_skb_gso(vif, skb, gso)) {
1287 /* Failure in xenvif_set_skb_gso is fatal. */
1293 /* XXX could copy straight to head */
1294 page = xenvif_alloc_page(vif, pending_idx);
1297 xenvif_tx_err(vif, &txreq, idx);
1301 gop->source.u.ref = txreq.gref;
1302 gop->source.domid = vif->domid;
1303 gop->source.offset = txreq.offset;
1305 gop->dest.u.gmfn = virt_to_mfn(page_address(page));
1306 gop->dest.domid = DOMID_SELF;
1307 gop->dest.offset = txreq.offset;
1309 gop->len = txreq.size;
1310 gop->flags = GNTCOPY_source_gref;
1314 memcpy(&vif->pending_tx_info[pending_idx].req,
1315 &txreq, sizeof(txreq));
1316 vif->pending_tx_info[pending_idx].head = index;
1317 *((u16 *)skb->data) = pending_idx;
1319 __skb_put(skb, data_len);
1321 skb_shinfo(skb)->nr_frags = ret;
1322 if (data_len < txreq.size) {
1323 skb_shinfo(skb)->nr_frags++;
1324 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1327 frag_set_pending_idx(&skb_shinfo(skb)->frags[0],
1328 INVALID_PENDING_IDX);
1331 vif->pending_cons++;
1333 request_gop = xenvif_get_requests(vif, skb, txfrags, gop);
1334 if (request_gop == NULL) {
1336 xenvif_tx_err(vif, &txreq, idx);
1341 __skb_queue_tail(&vif->tx_queue, skb);
1343 vif->tx.req_cons = idx;
1345 if ((gop-vif->tx_copy_ops) >= ARRAY_SIZE(vif->tx_copy_ops))
1349 return gop - vif->tx_copy_ops;
1353 static int xenvif_tx_submit(struct xenvif *vif, int budget)
1355 struct gnttab_copy *gop = vif->tx_copy_ops;
1356 struct sk_buff *skb;
1359 while (work_done < budget &&
1360 (skb = __skb_dequeue(&vif->tx_queue)) != NULL) {
1361 struct xen_netif_tx_request *txp;
1365 pending_idx = *((u16 *)skb->data);
1366 txp = &vif->pending_tx_info[pending_idx].req;
1368 /* Check the remap error code. */
1369 if (unlikely(xenvif_tx_check_gop(vif, skb, &gop))) {
1370 netdev_dbg(vif->dev, "netback grant failed.\n");
1371 skb_shinfo(skb)->nr_frags = 0;
1376 data_len = skb->len;
1378 (void *)(idx_to_kaddr(vif, pending_idx)|txp->offset),
1380 if (data_len < txp->size) {
1381 /* Append the packet payload as a fragment. */
1382 txp->offset += data_len;
1383 txp->size -= data_len;
1385 /* Schedule a response immediately. */
1386 xenvif_idx_release(vif, pending_idx,
1387 XEN_NETIF_RSP_OKAY);
1390 if (txp->flags & XEN_NETTXF_csum_blank)
1391 skb->ip_summed = CHECKSUM_PARTIAL;
1392 else if (txp->flags & XEN_NETTXF_data_validated)
1393 skb->ip_summed = CHECKSUM_UNNECESSARY;
1395 xenvif_fill_frags(vif, skb);
1398 * If the initial fragment was < PKT_PROT_LEN then
1399 * pull through some bytes from the other fragments to
1400 * increase the linear region to PKT_PROT_LEN bytes.
1402 if (skb_headlen(skb) < PKT_PROT_LEN && skb_is_nonlinear(skb)) {
1403 int target = min_t(int, skb->len, PKT_PROT_LEN);
1404 __pskb_pull_tail(skb, target - skb_headlen(skb));
1407 skb->dev = vif->dev;
1408 skb->protocol = eth_type_trans(skb, skb->dev);
1409 skb_reset_network_header(skb);
1411 if (checksum_setup(vif, skb)) {
1412 netdev_dbg(vif->dev,
1413 "Can't setup checksum in net_tx_action\n");
1418 skb_probe_transport_header(skb, 0);
1420 vif->dev->stats.rx_bytes += skb->len;
1421 vif->dev->stats.rx_packets++;
1425 netif_receive_skb(skb);
1431 /* Called after netfront has transmitted */
1432 int xenvif_tx_action(struct xenvif *vif, int budget)
1437 if (unlikely(!tx_work_todo(vif)))
1440 nr_gops = xenvif_tx_build_gops(vif);
1445 gnttab_batch_copy(vif->tx_copy_ops, nr_gops);
1447 work_done = xenvif_tx_submit(vif, nr_gops);
1452 static void xenvif_idx_release(struct xenvif *vif, u16 pending_idx,
1455 struct pending_tx_info *pending_tx_info;
1456 pending_ring_idx_t head;
1457 u16 peek; /* peek into next tx request */
1459 BUG_ON(vif->mmap_pages[pending_idx] == (void *)(~0UL));
1461 /* Already complete? */
1462 if (vif->mmap_pages[pending_idx] == NULL)
1465 pending_tx_info = &vif->pending_tx_info[pending_idx];
1467 head = pending_tx_info->head;
1469 BUG_ON(!pending_tx_is_head(vif, head));
1470 BUG_ON(vif->pending_ring[pending_index(head)] != pending_idx);
1473 pending_ring_idx_t index;
1474 pending_ring_idx_t idx = pending_index(head);
1475 u16 info_idx = vif->pending_ring[idx];
1477 pending_tx_info = &vif->pending_tx_info[info_idx];
1478 make_tx_response(vif, &pending_tx_info->req, status);
1480 /* Setting any number other than
1481 * INVALID_PENDING_RING_IDX indicates this slot is
1482 * starting a new packet / ending a previous packet.
1484 pending_tx_info->head = 0;
1486 index = pending_index(vif->pending_prod++);
1487 vif->pending_ring[index] = vif->pending_ring[info_idx];
1489 peek = vif->pending_ring[pending_index(++head)];
1491 } while (!pending_tx_is_head(vif, peek));
1493 put_page(vif->mmap_pages[pending_idx]);
1494 vif->mmap_pages[pending_idx] = NULL;
1498 static void make_tx_response(struct xenvif *vif,
1499 struct xen_netif_tx_request *txp,
1502 RING_IDX i = vif->tx.rsp_prod_pvt;
1503 struct xen_netif_tx_response *resp;
1506 resp = RING_GET_RESPONSE(&vif->tx, i);
1510 if (txp->flags & XEN_NETTXF_extra_info)
1511 RING_GET_RESPONSE(&vif->tx, ++i)->status = XEN_NETIF_RSP_NULL;
1513 vif->tx.rsp_prod_pvt = ++i;
1514 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&vif->tx, notify);
1516 notify_remote_via_irq(vif->tx_irq);
1519 static struct xen_netif_rx_response *make_rx_response(struct xenvif *vif,
1526 RING_IDX i = vif->rx.rsp_prod_pvt;
1527 struct xen_netif_rx_response *resp;
1529 resp = RING_GET_RESPONSE(&vif->rx, i);
1530 resp->offset = offset;
1531 resp->flags = flags;
1533 resp->status = (s16)size;
1535 resp->status = (s16)st;
1537 vif->rx.rsp_prod_pvt = ++i;
1542 static inline int rx_work_todo(struct xenvif *vif)
1544 return !skb_queue_empty(&vif->rx_queue);
1547 static inline int tx_work_todo(struct xenvif *vif)
1550 if (likely(RING_HAS_UNCONSUMED_REQUESTS(&vif->tx)) &&
1551 (nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
1552 < MAX_PENDING_REQS))
1558 void xenvif_unmap_frontend_rings(struct xenvif *vif)
1561 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1564 xenbus_unmap_ring_vfree(xenvif_to_xenbus_device(vif),
1568 int xenvif_map_frontend_rings(struct xenvif *vif,
1569 grant_ref_t tx_ring_ref,
1570 grant_ref_t rx_ring_ref)
1573 struct xen_netif_tx_sring *txs;
1574 struct xen_netif_rx_sring *rxs;
1578 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1579 tx_ring_ref, &addr);
1583 txs = (struct xen_netif_tx_sring *)addr;
1584 BACK_RING_INIT(&vif->tx, txs, PAGE_SIZE);
1586 err = xenbus_map_ring_valloc(xenvif_to_xenbus_device(vif),
1587 rx_ring_ref, &addr);
1591 rxs = (struct xen_netif_rx_sring *)addr;
1592 BACK_RING_INIT(&vif->rx, rxs, PAGE_SIZE);
1594 vif->rx_req_cons_peek = 0;
1599 xenvif_unmap_frontend_rings(vif);
1603 int xenvif_kthread(void *data)
1605 struct xenvif *vif = data;
1607 while (!kthread_should_stop()) {
1608 wait_event_interruptible(vif->wq,
1609 rx_work_todo(vif) ||
1610 kthread_should_stop());
1611 if (kthread_should_stop())
1614 if (rx_work_todo(vif))
1615 xenvif_rx_action(vif);
1623 static int __init netback_init(void)
1630 if (fatal_skb_slots < XEN_NETBK_LEGACY_SLOTS_MAX) {
1631 pr_info("fatal_skb_slots too small (%d), bump it to XEN_NETBK_LEGACY_SLOTS_MAX (%d)\n",
1632 fatal_skb_slots, XEN_NETBK_LEGACY_SLOTS_MAX);
1633 fatal_skb_slots = XEN_NETBK_LEGACY_SLOTS_MAX;
1636 rc = xenvif_xenbus_init();
1646 module_init(netback_init);
1648 static void __exit netback_fini(void)
1650 xenvif_xenbus_fini();
1652 module_exit(netback_fini);
1654 MODULE_LICENSE("Dual BSD/GPL");
1655 MODULE_ALIAS("xen-backend:vif");